Method for training a system adapted for reconstructing a medical visualising data package

Abstract

The invention is a training method for training a system adapted for reconstructing a medical visualising data package, wherein the system comprises a machine learning-based reconstruction module (200) and forward projection module (300), and, in course of the training method: - first pre-training (S220) and, after that, second pre-training (S330) are performed on the reconstruction module (200) and forward projection module (300) with respective loss functions, respectively, and - after that, first further training (S440) and second further training (S450) are performed on the reconstruction module (200) and forward projection module (300), respectively, with model reproduction loss function, by processing, by the forward projection module (300) and subsequently by the reconstruction module (200), a medical field-specific training set of volumetric visualising data packages generated based on a medical field-specific artificial investigation area model. The invention is furthermore a system for reconstructing a medical visualising data package.

Description

METHOD FOR TRAINING A SYSTEM ADAPTED FOR RECONSTRUCTING A MEDICAL VISUALISING DATA PACKAGETECHNICAL FIELDThe invention relates to a system adapted for reconstructing a medical visualising data package (for example a static image or a or time-varying image), and to a method adapted for training the system.BACKGROUND ARTImage reconstruction methods and the systems, apparatuses implementing them have become widely applied as tools in modem medical diagnostics. Thanks to the spreading of artificial intelligence, machine learning, and particularly of neural networks, several related applications have appeared in the fields of medical diagnostics and image reconstruction.Such an approach is disclosed in US 2022 / 0101526 A1 , wherein a method for generating a trained machine learning algorithm is disclosed. A deep learning-based medical imaging approach is disclosed in US 2018 / 0360313 A1 .An apparatus for processing medical information is disclosed in US 11 ,244,480 B2, and a medical imaging system having improved performance with the help of neural networks is disclosed in US 2022 / 0327665 A1 . According to US 2020 / 0027530 A1 , patient data are simulated on the basis of data of real patients, that is, models are generated for training the artificial intelligence algorithms.Approaches applying reconstruction are disclosed in certain articles (A. Mehranian et al.: Model-Based Deep Learning PET Image Reconstruction Using Forward- Backward Splitting Expectation-Maximization, IEEE Transactions on Radiation and Plasma Medical Sciences, Vol. 5, Issue: 1 , pp. 54-64, January 2021 ; K. Gong et al.: Iterative PET Image Reconstruction Using Convolutional Neural Network Representation, IEEE Trans Med Imaging, 38(3): 675-685, Mar 2019; J. Adler, O. Oktem: Learned Primer-dual Reconstruction, arXiv:1707.06474v3, 2018).Medical imaging typically involves medical images recorded by SPECT, PET, CT, MRI, optical imaging apparatus and ultrasonic apparatus. In accordance with theimaging apparatus, their modality is SPECT, PET, CT, MRI, optical and ultrasonic, respectively.In SPECT imaging (Single-photon emission computed tomography) the gamma camera head is rotated around the patient, by means of the imaging of the distribution of the gamma-emitting isotope injected into the patient is performed. The 3D (or, in the case of temporal processes, 4D) activity distribution can be calculated from the measured SPECT images applying suitable mathematical algorithms.In PET imaging (Positron emission tomography) the tracer material injected into the patient comprises an isotope undergoing a decay with positron emission, each emitting, through positron annihilation, two 511 keV gamma photons. The PET apparatus detects these photons in coincidence; the activity distribution can be calculated by suitable mathematical algorithms from the detected photon pairs.CT investigation (Computed tomography) involve irradiating the patient with X-ray; by measuring the absorption of X-ray photons in bodily tissues from various directions, applying a suitable mathematical algorithm it is possible to generate a tomography 3D image that is proportional to the electron density of the body and is characteristic of the patient’s tissue morphology.MRI investigation (Magnetic resonance imaging) involves placing the patient in a strong magnetic field, followed by exciting the magnetic moments of hydrogen atoms of the tissues with a radiofrequency signal. The electric signal emitted during the relaxation of the magnetic moments is detected; an image of the 3D hydrogen density distribution in the tissue, which is characteristic of the patient’s tissue morphology, can be reconstructed from the detected signal sequences applying suitable mathematical methods.A common characteristic of the above-mentioned measurement methods applicable for medical diagnostics and medical-biology research is that significant noise is present in addition to the detected signal. In case of SPECT, for example, the noise (photons are detected and this detection is burdened with noise) strongly depends on the activity injected into the patient, the measurement time, the absorbance of the body, and characteristics of the detector (e.g. gamma collimator configuration).In view of the known approaches, there is a need for a system adapted for reconstructing a medical visualising data package (for example, a medical image) that can be effectively applied for generating an appropriate reconstruction.DISCLOSURE OF THE INVENTIONThe primary object of the invention is to provide a system for reconstructing a medical visualising data package (for example a medical image) and a training method therefor which are free of the disadvantages of prior art approaches to the greatest possible extent.The object of the invention is to provide a system for reconstructing a medical visualising data package (for example a medical image) and a training method therefor, wherein the system can be applied for effectively producing, in a medical field-specific manner, preferably even involving a certain degree of personalisation (customisation), appropriate reconstructions that preferably have high signal-to- noise ratio. The personal customisation feature ensures that the reconstruction is based not only on the physical properties of imaging and on the results of previous measurements, but preferably also provides consistency with the current measurement for each patient examination, which is important for reliability. As it will be seen, the reconstruction may either be modality-preserving (modalitykeeping; i.e., the modality of the volumetric (volumetrical) visualising data package reconstructed by the reconstruction module is identical to the modality of the medical visualising data package from which it was generated), or modality-switching (modality-changing; the modality of the reconstructed volumetric visualising data package is different from the modality of the medical visualising data package fed to the reconstruction module).The invention can be regarded more effective than the known reconstruction approaches because instead of providing an approach with a wider, more general aim it offers a solution that is medical field specific (medical-application specific), and preferably also takes into consideration (i.e., learns from) the results of the specific (concrete) measurement. Thereby, by applying an appropriate training sequence and defining the given medical field (application) it could achieve better results.Because of subjecting the patient and the physician to a dose, an important object set before the invention is to reduce the activity introduced into the body of the patient under examination and to reduce the measurement time; the latter also reduces the probability of motion blurring, and increases the number of patients that can be examined with the apparatus, also improving the effectiveness and return on investment of the examination. The reduction of activity and measurement time both reduce proportionally the number of detected photons, and thus it reduces the signal-to-noise ratio, which in turn reduces the probability of safely recognition (identification) of accumulations being different (deviating) from normal, and thereby affects the diagnostic value of the measurement process. The invention preferably provides a solution path for preserving the diagnostic value of the examination method also in the case of reducing activity, dose.The above considerations can be referred to SPECT and similarly can be said also for PET. In the case of CT, the reduction of X-ray flux is a fundamental aim for reducing the dose the patient is exposed to, and thus for reducing the risks of the examination. With MRI, it is expedient to aim at the significant reduction of the recording time required for taking high-resolution images.The objects of the invention can be achieved by providing the training method according to claim 1 and the system according to claim 15. Preferred embodiments of the invention are defined in the dependent claims.The efforts described above in relation to the objects involve, in many cases, that the signal-to-noise ratio of the measured images may decrease, even significantly. An important characteristic of the present invention is that it preferably provides a solution for reconstructing the measurements, i.e. , according to the given modality,- for a distribution characteristic of the 3D morphology of a body tissue (for example CT and MRI), as well as- for a 3D or 4D distribution of a radiopharmaceutical proportional to the biological process being examined (for example SPECT and PET), which is characterised by a diagnostically acceptable (i.e., appropriately high) signal-to-noise ratio even if the measured information has low signal-to-noise ratio. This is achieved by applying a system including a plurality of neural networks that are first trained separately and subsequently together, and, in relation to that, analgorithm sequence based on multiple-stage (multiple-phase) machine learning (artificial intelligence).State-of-the-art image reconstruction methods take into consideration the physics of the imaging; for example, in the case of SPECT and PET, the processes of gamma ray absorption and scattering occurring in the patient’s body and in the structural material of the detector. This is preferably also included in the present invention. Physics can be taken into consideration using the Monte Carlo reconstruction with which, preferably, the training set is generated in stage 1 (phase 1 ). After that, i.e., in stages 2 and 3 (phases 2 and 3), training is performed applying these sets that contain implicitly, i.e., “built into” them the physics of imaging. It is noted that in stages 2 and 3 the effectiveness of learning can be improved by optionally applying such a neural network, which is a so-called “physics informed” system.In the course of image reconstruction (i.e., in general, here and elsewhere: reconstruction of visualising data packages) for determining the values of the elementary units of the volume to be reconstructed (e.g. voxels, but other representation may also be applied) a system of linear equations typically consisting of millions of equations must be solved approximatively. The most cutting-edge classical solutions for performing these calculations (i.e., solutions not based on artificial intelligence, that is, in the present case, on machine learning; such solutions are collectively referred to as different-from -machine learning-based) for example apply Monte Carlo methods for estimating the distribution of activity with the required accuracy. The computation cost of the typically iterative image reconstruction methods is, however, very high, and even when run on supercomputing architectures included several hundreds or even thousands of high-performance processors, e.g. GPU or GPU clusters, they take significant time to complete.This causes problems especially in case not only spatial (3D) information is of interest but, e.g. by dividing the measurement into short time slices, several 3D images (i.e., a 4D process) is to be reconstructed. In such a case, for k pieces of time slices the reconstruction time typically increases k-fold, with the signal-to-noise ratio of each time slice typically also becoming significantly deteriorated.An important and preferably appearing characteristic of the machine learning-based solution according to the invention is that, on the one hand it inherits the accuracy of the cutting-edge image reconstruction method (e.g. Monte Carlo) applying modelling (see the preferably applicable stage 1 ) as the two neural networks are pretrained in stages 2 and 3 based on them. On the other hand, its running time significantly reduces relative to these, even by one or two orders of magnitude, which considerably improves its applicability. This is achieved by training machine learning-based (typically, neural network-based) modules (computation units) capable of performing the computation, the process of which will be described in detail below.The applied image reconstruction methods are characterised by that their operation depends on several - typically hundreds of - reconstruction parameters (i.e., a plurality of reconstruction parameters correspond to a given reconstruction method, as well as a reconstruction module). Such reconstruction parameters are for example the number of iterations, the type and parameters of the internal regularisation (noise-suppressing) filter, in the case of Monte Carlo algorithms, the number of simulated photons applied, to which computation accuracy is proportional, and which is in direct proportion to the running time.The typical activity distribution and tissue density distribution of certain examination types strongly depend not only on the measurement device but also on the examination protocol (such as the injected activity per kg of bodyweight, and the type of the applied tracer material and isotope). In addition to that, it strongly depends on the patient, for example the absorbance (in the case of CT) or isotope uptake capability (in the technical terminology: uptake, for example in the case of SPECT or PET) of the given organ or tissue. It also depends on the body size and physiological specificities of the patient.The result of these is that the parameters of the reconstruction algorithms must be adjusted (optimised) for the given type of examination, in other words, for medical field (application) in order to achieve appropriate results. This, on the one hand, is a very time-consuming task, but, on the other, with the currently known technical approaches it is impossible to optimise the reconstruction for a given patient.An important further characteristic of the invention (which can also be called artificial intelligence-supported reconstruction of an image or visualising data package) is that it also implements reconstruction settings that are optimised for a given medical field (see stage 4 (phase 4) below), and preferably at the same time also for a given patient (see the optional stage 5 (phase 5) below). This is performed in such a manner that it take particular attention of the diagnostic accuracy of the biomarker that is to be measured in the course of quantitative imaging performed in the given medical field.Image processing and image reconstruction methods based on machine learning (ML, or, somewhat more loosely, on artificial intelligence - Al), are becoming more and more widespread (such methods are generally applicable also for visualising data packages that consist of separated images and are therefore time-dependent). A characteristic of these Al-based solutions is that the features included in the images (the so-called “feature vector”) and the structures (e.g. the shape or image features of the ribs and vertebrae) included therein are not determined “manually” (or in the jargon of the field: “in a hand-designed manner”) but are identified and selected by a self-learning process, for example, such an auto-representation is produced in the deep layers of convolutional neural networks (CNN, e.g. in autoencoder-type neural network architectures).Image reconstruction methods that apply Al appeared a few years ago, and are considered an intensively researched field. Several approaches concerning Al- based image reconstruction methods are known. Such a method is for example direct neural network-based reconstruction, wherein the input of the neural network is the noise-burdened measurement to be reconstructed, and its expected output is the reconstructed volumetric image (or simply: reconstructed volume) having high signal-to-noise ratio.It is known that in the case of the method termed the “direct solution”, training requires an extensively large number of data (pair of measured and reconstructed images) in order to exceed the performance of traditional reconstruction methods (such as for example, Monte Carlo reconstruction). A good example for this is DPIR- Net (Zhanli Hu, et al.: DPIR-Net: Direct PET Image Reconstruction Based on the Wasserstein Generative Adversial Network, IEEE transactions on radiation andplasma medical sciences, vol. 5, no. 1 , p. 35-43, January 2021 ), where nearly 38,000 images were utilised for training the neural network to achieve the expected performance.Obtaining such a large number of medical images presents great difficulties (e.g. problems related to ownership rights, ethical problems). On the other hand, if the clinical protocol is modified and the information content and the properties of the measurements are changed, according to the known methods the already trained neural network must be re-trained or trained further, and the required clinical verification must also be performed, which typically involves costs that are prohibitive for the practical application of this type of approaches.It is known, however, that the size of the training database can be reduced to a few dozen images in case for example images filtered by artificial intelligence are utilised for the regularisation of reconstruction, for example by applying a GAN (Generative Adversarial Network; Zhaoheng Xie et al.: Generative adversarial network based regularized image reconstruction for PET, Phys. Med. Biol. 65, 125016, 2020); this is because in this framework the physical model inherently included in the reconstruction algorithm is preserved, which helps the convergence of the reconstruction, but the computation still requires significant time and resources. Thus, this reconstruction method does not yield significant progress in speed with respect to the prior art high-accuracy approaches (applying e.g. Monte Carlo methods).A method called “conditional deep image prior framework” is also in use, wherein a neural network with suitable topology is trained, applying data of a larger population, such that it learns characteristic structures and is able to synthesise them. Subsequently, the network further learns in-situ (i.e. at the location of the examination) with the help of the given measurement (see Jianan Cui et al.: Population and individual information guided PET image denoising using deep neural network, Conference Paper, May 2019, Conference: The Fifteenth International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine), such that the reconstructed image applies not only the structures trained on population data but also the information contained in the measurements related to specific, individual patients.These are e.g. applied for training an appropriate autoencoder to synthesising a PET image on the basis of CT images. Because the CT image does not contain the physiologically relevant information, the CT image only kind of addresses the potential structures in the reduced latent space (i.e., in the deep layers of the network where the structures characteristic of the population are encoded), synthesisation being made more accurate by an in-situ training stage utilising a specific measurement (see in the article related to the above by Jianan Cui et al.: PET image denoising using unsupervised deep learning, European Journal of Nuclear Medicine and Molecular Imaging, 2019 Dec; 46(13): 2780-2789). Modality switching is not applied, because here the CT image is utilised for addressing the most appropriate one of the many possible PET images. The CT image is treated as additional information, but a PET image is not synthesised from the CT image.It is emphasized that the invention provides extremely great advantage with respect to prior art approaches thanks to its training structure based on consecutive training stages (for the stages applied in the invention and some optional stages see the description included below).An important characteristic of the solution according to the invention is that it is not necessary to have a large number of measurements for the appropriate training of the network; it uses both information being in the population and individual information such that the resulting reconstruction (reconstructed image, in general: reconstructed visualising data package) has such features that it realises a reconstruction that is optimal in regards of the given medical field-specific measurement numbers, that is, reconstruction accuracy and the diagnostic value of these measurement numbers are automatically maximised by it (thanks to the medical field-specific training step, see below). This characteristic makes the implementation according to the invention largely invariant to the change of the measurement protocol (this invariance or flexibility is an important characteristic of neural networks, and precisely this characteristic is emphasized by the applied special training order), so the practical applicability of the method and the apparatus implementing the method is significantly improved.It is important to also emphasise that a neural network or a system consisting of neural networks is determined by three things: a) the structure of the network andthe relative position of the networks; b) the characteristics and process of training; c) the training database. Difference in any of the components - considering that there are strongly non-linear systems are in question - may result in significant differences in the end result and the functionality of the system.In the invention, therefore, a specific training scheme is applied, and the entirety of the reconstruction operator is replaced by a system of artificial intelligence-based modules (e.g. neural networks), and the training is carried out such that it is optimal from the aspect of the given medical field. In the case of the invention, therefore, it is ensured by a well-defined training sequence and a system based on artificial intelligence (e.g., neural networks) that the system can accurately generate the features diagnostically valuable for the chosen specific medical field based on a relatively low amount of training data (images or, in general, visualising data packages).BRIEF DESCRIPTION OF THE DRAWINGSPreferred embodiments of the invention are described below by way of example with reference to the following drawings, whereFig. 1 illustrates in a flow chart the steps preferably applicable for generating a training database for the training method according to the invention in a (modality-preserving) embodiment (optional stage 1 ),Fig. 2 illustrates in a flow chart the pre-training of the reconstruction module in a (modality-preserving) embodiment of the training method according to the invention (stage 2),Fig. 3 illustrates in a flow chart the pre-training of the forward projection module in a (modality-preserving) embodiment of the training method according to the invention (stage 3),Fig. 4 illustrates in a flow chart the further training of the reconstruction and forward projection modules in a medical field-specific training step of the training method according to the invention in a (modality-preserving) embodiment (stage 4),Fig. 5 illustrates in a flow chart the further training of the reconstruction and forward projection modules in an optional post-training step of the trainingmethod according to the invention in a (modality-preserving) embodiment (optional stage 5),Fig. 6A is a schematic flow chart summarising the embodiment of the invention according to Figs. 1 -5,Fig. 6B illustrates a Bloch spere, andFigs. 7-11 illustrate in flow charts similar to Figs. 1 -5 the optional stage 1 , stages 2-4, and the also optional stage 5 in a modality-switching embodiment.MODES FOR CARRYING OUT THE INVENTIONThe invention relates to a training method for training a system adapted for reconstructing a medical visualising data package (for example a static image or a or time-varying image), wherein- the system comprises a reconstruction module based on machine learning, adapted for generating a volumetric visualising data package by processing a medical visualising data package (in the embodiment illustrated in Figs. 2-5, see a reconstruction module 200, with other name, reconstruction unit), and- such a system is applied in course of the training method (i.e. , this is also needed for performing the training method, after which it can be removed) which comprises a forward projection module based on machine learning, adapted for generating a medical visualising data package by processing a volumetric visualising data package (in the embodiment illustrated in Figs. 2-5, see a forward projection module 300, with other name, forward projection unit).In an embodiment, therefore, the visualising data package is a static image or a time-varying image (image changing over time; this holds true for all kinds of medical and volumetric visualising data packages).In the following, the invention will be defined in general, and a further-specified embodiment is described below that in relation to the figures. In the illustrated embodiment, the invention is preferably complemented with a stage 1 related to preparing a training database (applying a splitter module and / or reconstruction parameters, in relation to this see Figs. 1 and 7; there are differences also in the preparation process of these). However, the training database may also be prepared in a different way, i.e., a generic training database may also be applied that can betreated as given (see the general introduction to this invention below wherein no constraints are included in relation to preparing the training database).The training method according to the invention comprises the following steps (that is, the following steps are applied in the course of the method according to the invention):- In a first pre-training step (in the embodiment illustrated in the figures, this is the so-called stage 2 disclosed by the help of Fig. 2, so references will be made to this figure) a first pre-training volumetric visualising data package (see e.g. a first pre-training volumetric image 210 in Fig. 2, as an exemplary first pre-training volumetric visualising data package) is generated for each of those (i.e., a respective for each visualising data package of the set) by processing a first auxiliary pre-training set of medical visualising data packages (see e.g. the medical images of a first auxiliary pre-training set 115 in Fig. 2, as medical visualising data packages of an exemplary first auxiliary pre-training set) by means of the reconstruction module, and performing a first pre-training (see in Fig. 2 operational step S220 corresponding to the first pre-training) on the reconstruction module (in other words, a first pre-training of the reconstruction module is performed; these formulations related to training can also be used in relation to other training operations and modules), with a first pre-training loss function determined (generated) based on (or, in other words, in the case of such constructions: based on their comparison) a second auxiliary pre-training set of volumetric visualising data packages (see e.g. a second auxiliary pre-training set 120 of volumetric images in Fig. 2 as volumetric visualising data packages of an exemplary second auxiliary pre-training set) corresponding to the first auxiliary pre-training set of the medical visualising data packages and on the first pretraining volumetric visualising data packages.- In a second pre-training step (in the embodiment according to the figures this is the so-called stage 3 that is illustrated in Fig. 3, so references will be made to this figure; the first and second pre-trainings are performed one after the other since the reconstruction module is to be trained first) after the first pre-training step a second pre-training volumetric visualising data package is generated for each of those (see e.g. in Fig. 3 a second pre-training volumetric image 310 asan exemplary second pre-training volumetric visualising data package; it is marked differently from the first pre-training volumetric image 210 applied in stage 2 because the volumetric image 310 is generated by means of the reconstruction module 200 that has already been pre-trained) by processing the first auxiliary pre-training sets of the medical visualising data packages (see Fig.3, by way of example the medical images of the first auxiliary pre-training set 115 are applied also in this step) by means of the reconstruction module, a second pre-training medical visualising data package (see e.g. second pre-training medical images 320 in Fig. 3 as exemplary second pre-training medical visualising data packages) is generated for each of those by processing each of the second pre-training volumetric visualising data packages by means of the forward projection module, and a second pre-training (see in Fig. 3 an operational step S330 corresponding to the second pre-training) is performed on the forward projection module with a second pre-training loss function determined based on visualisation information corresponding to the medical visualising data packages of the first auxiliary pre-training set of the medical visualising data packages and on the second pre-training medical visualising data packages.- In a medical field-specific training step (in the embodiment illustrated in the figures, this is the so-called stage 4 illustrated by Fig. 4, so references will be made to this figure) after the second pre-training step a medical field-specific training medical visualising data package (see e.g. a medical image 420 in Fig.4, as an exemplary medical field-specific training medical visualising data package) is generated for each of those by processing a medical field-specific training set of volumetric visualising data packages (see e.g. volumetric images of the medical field-specific training set 410 in Fig. 4 as the volumetric visualising data packages of an exemplary medical field-specific training set; the terms “first and second auxiliary pre-training set” are applied because these are sets applied in a two-stage pre-training, the medical field-specific training set is applied in the medical field-specific training step) generated based on a medical field-specific artificial investigation area model by means of the forward projection module, medical field-specific training loss-check volumetric visualising data packages (see e.g. medical field-specific training loss-check volumetric images 430 in Fig.4 as exemplary medical field-specific training loss-check volumetric visualising data packages) are generated by processing the medical field-specific training medical visualising data packages by means of the reconstruction module, and a first further training and a second further training (see in Fig. 4 respective operational steps S440, S450 corresponding to the medical field-specific training step) are performed on the reconstruction module and the forward projection module, respectively, with a model reproduction loss function determined based on the medical field-specific training set of the volumetric visualising data packages and the medical field-specific training loss-check volumetric visualising data packages.Typically, in loss functions corresponding elements are compared pair by pair (for example by difference-forming, see for example the losses L1 , L2 in stages 2 and 3, or the interpretation of the formula disclosed in stage 4, as well as such can also be applied in stage 5), with this introduction it is intended to show that these are the inputs of the loss function which can be formulated in many ways (which are connected by the phrase “based on”).It is noted that the visualisation information - which is image information in an embodiment, as described below, but in accordance with the applied terminology it is generalised to visualising data package - can be a medical image of the first auxiliary pre-training set itself, for example a medical image of a first auxiliary pretraining set 115, but it can also be a medical image from a base set 110, such as in Fig. 3. In the latter case, the members (elements) of the first auxiliary pre-training set 115 are generated from members of the base set 110, so in this case the visualisation information thereof is considered to correspond to respective medical visualising data packages of the first auxiliary pre-training set (the same element of the base set corresponds to multiple indices j, see the mode of generating of the first auxiliary pre-training set 115). In accordance with the generating, the manner of the correspondence can be established, see also below in the description of the respective embodiment. It can be also said that the visualisation information is constituted either by medical visualising data packages of the first auxiliary pretraining set or the medical visualising data packages they originate from.Above, the adjective “loss-check” can also be omitted, it has been included for differentiating in the term designating the visualising data package applied for the loss function when the mapping “completes a cycle”, i.e. in this case, starting from a volumetric visualising data package a volumetric visualising data package has again been obtained by applying the modules. The adjective “loss-check” used for other visualising data packages could also be used for the second pre-training medical visualising data package, i.e., it could also be called a second pre-training loss-check medical visualising data package; however, in this case, the name is easily distinguishable. Instead of the adjective “loss-check”, the adjective “derived” could also be used at the appropriate places.In stage 4, therefore, medical field-specific training is performed with the help of a medical field-specific artificial investigation area model (the name of the model is shortly: artificial model). The term specific for medical field (in other words, medical application specific) is taken to mean specificity to a given specialist field, medicinal field, in other words, specificity to a given type / kind of medical examination (i.e., that the potential presence of which type of abnormality is to be visualised and examined; see also the related illustrative examples). Accordingly, stage 4 prepares the reconstruction module and the forward projection module to that what kind of abnormality and alteration are to be identified in the given medical field, i.e., what kind of abnormality or alteration may be experienced. To this - in line with the usual training principles - the training database, i.e., the artificial investigation area model will comprise also abnormal and healthy, slightly abnormal, as well as extensively abnormal training data, such that the trained modules are prepared appropriately (cf. the possible definition of the loss function disclosed in relation to stage 4). The artificial investigation area model is specified by an appropriate framework system (typically implemented by software on a computer), and the required training data can be generated based on this specification of the model. Accordingly, these training data belong to the model. Therefore, the artificial investigation area model models an area of investigation (region of examination; typically, of the body), which has unique configuration details (e.g. fitting to bodily constitution, bodily dimensions, etc.) and may exhibit abnormality. The framework of the artificial investigation area model specifies the details of the configuration details that can correspond to certain subregions.The artificial investigation area model could also be called an artificial medical investigation area model; however, we have not chosen the latter term because the Hungarian language does not differentiate in the term “area” from the “medical fieldspecific” concept. It is noted that the artificial investigation area model can also be called an artificial investigation area model of the examined region of a body (i.e., a human or animal body). It is an important consideration that in stage 4 training is performed applying visualising data packages, i.e., for example volumetric images, generated with this model, while in stages 2 and 3 real images are utilised for training.As it will be described in detail below, in the training method according to the invention in-situ post-training is preferably also applied prior to generating the reconstructed volumetric visualising data package (for example, image) that functions as the output of the system. This training sub-process is referred to below as stage 5, and an exemplary implementation is illustrated in Fig. 5 (in relation to which see also further below).Accordingly, in an embodiment of the training method according to the invention (within this, a further-constrained embodiment is illustrated in Fig. 5 illustrating stage 5), in a post-training step after performing the medical field-specific training step,- a post-training volumetric visualising data package (see e.g. in Fig. 5 a posttraining volumetric image 530 as an exemplary post-training volumetric visualising data package) is generated by processing a post-training input medical visualising data package (e.g., this is an actually measured medical image, i.e., a measurement medical image, that is, preferably a medical visualising data package originating from a local / in-situ measurement, see e.g. in Fig. 5 a post-training input medical image 500 as an exemplary post-training input medical visualising data package) by means of the reconstruction module,- a post-training loss-check medical visualising data package (see e.g. in Fig 5 a post-training loss-check medical image 535 as an exemplary post-training losscheck medical visualising data package; the adjective “loss-check” can also be omitted, it has been included in the term for differentiating the visualising data package used for the loss function when the mapping “completes a cycle”, i.e. in such a case, starting from a medical image a medical image has again beenobtained by applying the modules) is generated by processing the post-training volumetric visualising data package with the forward projection module, and- a third further training (see operational step S540 in Fig. 5) and a fourth further training (see operational step S550 in Fig. 5) is performed on the reconstruction module and on the forward projection module, respectively, with a reproductionappropriateness loss function determined based on the post-training input medical visualising data package and the post-training loss-check medical visualising data package.Preferably therefore, after performing the medical field-specific training step, another post-training step is performed as described above. An important specific feature of post-training is that it is carried out after medical field-specific training, i.e., on a system that has already been trained for a specific medical field, of course expediently applying such visualising data packages that can be linked to the specific medical field. Thus, stage 5 constitutes a type of back-check in order to ensure consistency. Furthermore, to ensure that the output is affected not only by population data (i.e. previous training) but also by the current measurement. Training can also be performed without stage 5, but in certain cases performing stage 5 may have a very large effect on the output due to back-checking, so to provide uniform-consistency for the results it is expedient to also include posttraining. The steps above - which constitute an iteration cycle - may also be performed in sequence more than once, i.e., more than one iteration cycles can be performed for post-training the reconstruction module and the forward projection module based on the input medical image (which has been typically recorded in- sitw, see the discussion of the iteration cycle herebelow).It is noted here that the term “loss” is widely used in this technical field to cover, as a simplified term, loss functions to be fed back into the machine learning modules during training, or - as shown by the context - in certain cases, for a given single data element of the loss function.In stage 5, various types of loss may be worth using as a reproductionappropriateness (reconstruction-appropriateness; reproduction-adequateness, reproduction-properness) loss function (or, in an alternative term, reproductiongoodness; of these the likelihood-based functions are referred to, for short, as“likelihood loss"), it can even be L1 or L2 loss (and also an adaptive loss function: the loss function also includes parameters that are set by neural network; see Jonathan T. Barron: Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 4331 -4339), but preferred choices may be among the loss functions (each providing a separate embodiment), if the reproduction-appropriateness loss function is- a Poisson-based maximum log-likelihood (logarithm-likelihood) loss function (C. Gourieroux, A. Monfort, A. Trognon, Pseudo Maximum Likelihood Methods: Applications to Poisson Models, Econometrica, Vol. 52, No. 3 (May, 1984), pp. 701 -720), or- a structural similarity based loss function (or in other words, a structural similarity index / index measure (SSIM) based loss function, see Mudeng V, Kim M, Choe S-w. Prospects of Structural Similarity Index for Medical Image Analysis. Applied Sciences. 2022; 12(8):3754; see also H. Zhao et al.: Loss Functions for Image Restoration with Neural Networks, arXiv: 1511 .08861 v3, 2018).In both cases, therefore, a function can be defined as the loss function, for which an extreme value (extremum) is searched for (in line with the loss-approach, a minimum is typically searched for). Also, in the case of a loss function based on structural similarity, a scalar function is defined which depends on the state of each neuron and on the training images. That is, the loss - be it any type of representation- defines the multi-dimensional configuration space together with the training database and the neuron-structure.It is noted that in the model reproduction (reconstruction) loss function in stage 4 and the reproduction-appropriateness loss function in stage 5 can also be called, respectively, a medical field-specific loss function and a post-training loss function.It is also noted that the post-training step can be considered a separate post-training method (even then, in case a “back-measurement” to stage 4, or intermediate application thereof is included in that, see below) that is performed on an appropriately trained system. This method is typically performed at the installation location, in an online manner, i.e. such that the output is also generated shortly after the post-training is performed applying the actual measurements (post-training, especially in a quantum representation, can be carried out quickly, which providesan advantage in connection with the local quick reaction). Post-training - in case it is to be carried out according to the embodiment - is expediently configured such that no user intervention is required for performing thereof, but post-training is performed in line with the steps of the method and the configuration of the system. That is, if the system receives a medical image measured in-situ as an input (more generally: a visualising data package), then it performs the steps of post-training applying that.In the present application the concepts (notions) of “medical image” (essentially, an image to be reconstructed, or in general, a visualising data package to be reconstructed) and “volumetric image” are applied (in a generalised manner, the terms “medical visualising data package” and “volumetric visualising data package” are applied; the generalisation can be made anywhere except if there is an explicit hindrance, see further below). These can correspond to the input and output sides of the reconstruction module (the reconstruction module performs the reconstruction of a medical image, i.e. , the volumetric image it outputs is the reconstruction of the input medical image, that is, in this case it is the reconstructed medical image), when a volumetric image is generated from a medical image by reconstruction. At the same time, these correspond also to the input and output sides of the forward projection module. In this case, the initial data is the volumetric image that is not necessarily the result of a reconstruction (cf. stage 4, wherein a volumetric image is the training data), that is why, in the case of the medical image and the volumetric image the terms “image to be reconstructed” and “reconstructed image” cannot be used in all cases (the former term is better usable). Interrelated medical images and volumetric images - which are generated from each other by the reconstruction module or the forward projection module - can be also referred to as “image pairs” (in the case of training data, even as “measured and reconstructed image pair”). Naturally, the information content of the medical image is such that it is possible to generate from it a reconstructed volumetric image applying the reconstruction module. The starting (input) image (in general: visualising data package) is termed a medical image because this is typically the image that is obtained as a medical recording (image) by the imaging / measurement according to the different modalities.The concept of “visualising (visualisation) data package”, which is more general than the concept of “image” has been introduced in order to be able to also take into account the time dependency of the data appearing in the invention. Therefore, the visualising data package can be a static image but also a time-varying image. And, as far as the volumetric visualising data package is concerned, it can either be a static volumetric image (3D) or a time-varying volumetric image (which can also be called a 4D image). Of course, the time dependency must be treated uniformly: if, for example, time-dependent training data are applied in stage 4, then such data have to also be applied in the training processes of the other stages. However, in the detailed description we essentially refer to images as visualising data packages; at appropriate places also referring to the aspects related to the extension to 4D. Furthermore, in the digital case, the image is represented for example by pixels and the volumetric image by voxels (the information content of pixels and voxels may change in case of time dependency).Certain embodiments of the invention relate to a system adapted for reconstructing a medical visualising data package (e.g. a medical image). The system according to the invention comprises a reconstruction module based on machine learning, adapted for generating a reconstructed volumetric image - constituting the output of the system - by processing a medical image, and it is trained by means of an embodiment of the training method according to the invention. The embodiments related to the system can also be considered as the module or modules of the system forming parts of an apparatus. Therefore, the embodiments related to the system can also be considered to relate to an apparatus adapted for reconstructing a medical visualising data package, and the training method to be adapted for training the apparatus.In an embodiment, the system according to the invention further comprises a forward projection module based on machine learning, adapted for generating a medical visualising data package by processing a reconstructed volumetric visualising data package, and in this embodiment the system is trained performing the training method - also applying the optional stage 5, i.e., post-training performed in-situ - set forth above. If the post-training stage, i.e., the optional stage 5 is notapplied, then the forward projection module can be removed from the trained system (i.e. , it can be removed after performing the training method).Furthermore, in accordance with the above, the invention can be considered to also relate to an apparatus (system) for reconstructing a medical image (in general: visualising data package) recorded by medical imaging apparatus, i.e., typically SPECT, PET, CT, MRI, or possibly by optical imaging and ultrasound equipment, obtained through various examinations (preclinical, clinical / medical diagnostic measurement methods and examinations, or measurements and examinations targeted at medical-biological research), as well as to the application of the system, i.e., to a reconstruction method performed by the trained reconstruction module.The effectiveness, accuracy, and processing speed of neural networks can be further improved by replacing conventional neurons with quantum neurons, i.e., for example by applying a quantum neural network (QNN), summarized, a quantum representation, in the reconstruction (see Jeswal, S.K., Chakraverty, S.: Recent Developments and Applications in Quantum Neural Network: A Review. Arch Computat Methods Eng 26, 793-807 (2019), and see also herebelow the possibilities related to that).In the case of quantum neural networks, the input images (generally, the visualising data packages) of the modules (the reconstruction module 200 and forward projection module 300) preferably implemented with neural networks are encoded in so-called qubits. In contrast to the conventional binary data representation, a qubit denotes a two-state quantum system, wherein the possible mixed states can be illustrated to the surface of a unit sphere with complex value (Bloch-sphere), see the diagram shown in Fig. 6B.In an embodiment, therefore, the medical visualising data packages (for example, static medical images or time-varying images) and the volumetric visualising data packages (for example, static volumetric images or time-varying volumetric images) are stored in qubits and a quantum representation is applied in the system.In case for example a measured image is encoded in quantum states, then preferably a significant data compression is also performed. For example in the caseof a NEQR (novel enhanced quantum representation) type quantum image representation, a greyscale image consisting of 1024*1024 pixel with a 8 bit / pixel representation can be encoded in as many as 28 qubits (1024=2A10; 10+10+8=28, i.e. two times 10 corresponding to the number of pixels plus 8 corresponding to the 8 bits, see in more detail in Wang, Z., Xu, M. & Zhang, Y.: Review of Quantum Image Processing. Arch. Computat. Methods Eng., 29(2), 737-761 (2022)).In case a quantum representation is applied in the present invention, the image reconstruction operation is performed, on an image encoded in qubits, by an appropriate quantum neural network, or by a network consisting of quantum gates of which the free parameters are set by a neural network (see below), the output of the operation being a (time-dependent) quantum state representing a 3D image (or a 4D one, for which the following description can be generalised).In an embodiment of the invention, therefore, quantum representation is applied in the system, that is, the medical and volumetric visualising data packages and the reconstruction and forward projection modules - in a trained system, optionally only the reconstruction module - are implemented in a quantum representation.Preferably, two solutions can be applied for implementing a quantum representation: full quantum representation and hybrid quantum representation.1. In the full quantum representation, preferably neural networks implemented by quantum neurons are applied in the reconstruction and forward projection modules, which directly acts on the bra and ket states representing the visualising data package (for example images), and is represented by qubits, i.e., the entire system is a quantum system having qubits as memory elements and quantum gates as operational elements applied to the qubits.2. In the hybrid quantum representation, the quantum gates are in a certain way “parametrised” by the neurons, i.e., nonlinear outputs of a classical neural network (this representation thus applies quantum gates based on classical neural networks), which quantum gates are then able to have effect on the bra and ket states.Such a hybrid quantum representation can be provided for example such that quantum gates are in a certain way parametrised by the neurons, that is, by the non-linear outputs of a classical neural network. For example, these adjust the degree of entanglement between two qubits, or the way of rotating the phase of a given qubit (these are necessary for a quantum-based implementation; based on the concept of entanglement introduced by the quantum approach, the states of objects located spatially further from each other are not independent from each other). The output is the measurement at the end of the quantum system. This is compared with the expected, and then the parameters of the quantum gates are adjusted so that to be appropriate. That is, the parameters of the parametrised network are considered as the parameters of a neural network, and these numbers are adjusted based on the feedback received from the loss function until the desired image (i.e. , an expected value thereof) appears at the output in an expected value.Both ways can be preferably applied in the invention, differing from each other essentially in the details of loss calculation, which will be addressed below.The relationship between the quantum state |p> of the measured (e.g. SPECT) projection image p and the activity distribution V, as well as its quantum state | \ / > is provided by the linear projection operator U:\p> = U \V>The reconstruction problem lies in calculating the components of, or inverting, the projection operator U or the system matrix. In the classical case, directly inverting the system matrix is impossible due to its large size. In the case of quantum states, the operator II is essentially a projector, i.e.:U = \p> <V| and, consequently, it has considerably smaller size than in the case of a conventional data representation. Inverting is required because inverting the system matrix is itself the solving the reconstruction problem. However, in the classical case this way is not feasible for many reasons. If we know the components of II, then we also know the reconstruction. We also known the reconstruction if we know the components of its inverse, because the base matrix and its inverse contain the same information. In the quantum case, it can be seen that the quantum representation may provide a solution, because II can be treated more “naturally” as it is a projector and it can be said that it is “delivered”.In an example, with a matrix size of 256*256, in the case of a SPECT recording comprising 128 projections and a 8-bit greyscale image, the state |p> can be represented in 8+8+7+8=31 qubits (NEQR quantum representation; 8+8 according to 256=28, 7 because 27=128, and 8 according to the 8 bits, see also the article referenced above). Similarly, let the volume |V> have 256*256*256 voxels, each voxel being represented in 10 bits, so in sum the state |V> can be represented in 34 qubits (8+8+8+10=34, see the explanation above). For comparison, IBM’s Eagle quantum computer made in 2021 has 127 qubits, while the Osprey system foreseen for 2022 will have 433, and the Condor system foreseen for 2023 will have 1121 qubits. Therefore, considering the number of necessary qubits, the reconstruction problem can be solved by the help of the available real quantum computing devices.Considering k known image pairs (i.e. , the index k is applied for indexing the images of the training database) as the paired elements of the training database (containing images p and the corresponding volumes V) the reconstruction problem can be formulated as follows (i.e., as the difference of the ket-bra projection operator |pk >< Vk| and the operator II): argminwhere fj runs through the free parameter values of the operator U: in a neural network representation these are the complex weights, threshold values and other free parameters of the quantum neurons; the state |xyO> and |xyzO> being the bases of the eigenvector (of the quantum state) of the projection image and the volume, respectively (the former has two spatial coordinates while the latter has three, i.e., these provide a basis for the images - considered as states - and accordingly depend on spatial coordinates). The formula is to be compared with the definition of the operator II above as a projection operator, according to which operator Up is searched for, what is - measured here in L1 norm - identical to solving the reconstruction problem. This can be checked by performing measurements on the bases. In other words, the operator II it is to be found that is closest to the projection operator composed of the image pairs pk, Vk. The eigenvectors |xyO> and |xyzO> are the bases (i.e., the image representation base system) of the quantum system representing the projection (medical) image and the volumetric image (in general, these are visualising data packages) that consists of qubits. In the case ofconventional images consisting of pixels, the location and value of the pixels span the pixelated image space. |xy0> and |xyz0> are elements of a Hilbert space, i.e. , elements of the space of functions.When, before performing the argmin operation, composing of the expected value is performed, a measurement is performed in the above-mentioned base system. By this measurement we can evaluate operators II corresponding to different parameters p, i.e., the argmin operation is effected on such evaluations. This involves repeatedly performing measurements on the quantum system according to the set of parameters p applied for adjusting the quantum operator, with the intention that this is to be fitted best to the training data. In the formula, adjusting the parameter p modifies the quantum operation before the measurement rather than after it. This is followed by the measurement <|> and then the comparison (subtraction of the two terms), which is followed by adjusting the value of p by the argmin operation in so that the difference is reduced.It is the inference process (see e.g. “deep learning inference”, i.e., execution) by the help of the projector (already “trained”) II established during training the quantum neural network with which a reconstructed volume corresponding to the medical image |p> is obtained:V = <xyzo| U* |p> where II* is the adjoint of II (thus, also in accordance with the formula above, the reconstruction of the state |p> is obtained by II*), i.e., the bra vector xyzO acts on the forward projected quantum state. In other words, after the operation performed with LT, a volumetric image represented in qubits is obtained. A measurement performed on this which projects this onto the xyzO base.In respect of the present invention, the great advantage of applying a quantum neural network is that, in a way, the quantum neural network is theoretically able to simultaneously store, in a latent space (which in this case is realised not by means of neuronal weights but by quantum states; the weights of the neural network adjust the operation of the quantum operator but the image itself is not realised in the neuronal weights but in a state determined by the qubits), all of the potential reconstructed images that are for example associated with different reconstruction settings. On the other hand, there are such methods available for training quantumneural networks which are able to find global extreme values more effectively than conventional gradient-based search methods (e.g. quantum annealing, see for example Arnab Das and B. Chakrabarti: Quantum annealing and related optimization methods, 2005, Springer-Verlag Berlin / Heidelberg), which can significantly speed up the training process.In the case of a quantum representation (quantum approach), in case of applying stage 5, the representation that is relevant for the given medical field (application) - i.e. , the appropriate reconstruction - is selected at the end of the multi-stage process (established by sequentially training more neural networks) by a specific measurement (see below in more detail in connection with the quantum representation of the optional stage 5, where the vq-outresult obtained by measurement, i.e., projection to a given base).Thanks to the applied training method, the resulting image generated in such a manner carries not only latent information from previous measurements, but realises an image reconstruction that is optimal from the aspect of the medical field, which additionally also takes into consideration the specialties of the specific measurement. The process can be verified on a quantum neural network simulator (e.g. Qibo, an open-source middleware / mediation software for quantum computations, https: / / qibo. science / ), which is able to run the given image reconstruction algorithm on both conventional computers and quantum computers.In line with the above (and see also further below), from the aspect of the present invention, its implementation on the quantum representation constitutes an especially preferred embodiment.To improve the accuracy of SPECT and PET image reconstruction, anatomy information provided by CT and MRI apparatus (for short: CT / MRI anatomy information associated with a SPECT / PET medical image) can be preferably applied for calculating the absorption properties of the body and the correction of gamma photon scattering (see below an exemplary mode of such use).In addition to that, the anatomy information provided by CT and MRI may contain common structural information. For example, in a bone-specific SPECT examinationthe isotope accumulates in the bone tissue, and thus CT images having strong contrast between bone tissue and soft tissue may also help the reconstruction by providing a priori information, as according to the expectations the tracer have to accumulate in the bone.Such a multi-channel neural network can be designed which is able to integrate this common structural information in an inherent manner, and thus further improve image quality (Zhaoheng Xie et al.: Anatomically aided PET image reconstruction using deep neural networks, Medical Physics 2021 Sep; 48(9): 5244-5258). Preferably, such network is applied in stages 2-5; see below in the description of these.However, in the case of the known solutions a large number of training images are still required for training such reconstructions, and it is also not capable of searching an optimum in the parameter space of the reconstruction operator. With the solution according to the invention such a reconstruction can be realised wherein all the features described in detail above are included at once as advantageous features (see below; reference is also made to introducing CT anatomy information in stage 1 and in stages 2-5, see the related Figs. 1 -5).As it is also illustrated in the figures, in this application a modality-preserving (Figs. 1 -5) and a modality-switching (Figs. 7-11 ) embodiment of the invention are described. The similarity of the figures shows that the two embodiments are based on the same principles (prescriptions in connection with the invention are introduced above), and thus after describing in detail the modality-preserving embodiment with the help of Figs. 1 -5, the modality-switching embodiment will only be discussed more briefly. Accordingly, in the embodiment of the method according to Figs. 1 -5, such a system is applied which is adapted for a modality-preserving reconstruction of a medical visualising data package (for example, a medical image), and the reconstruction module is configured (implemented) as a modality-preserving reconstruction module adapted for generating a modality-preserving volumetric visualising data package (for example, volumetric image) by processing a medical image (as it will be shown, Fig. 1 -5 in particular illustrate an embodiment that is modality-preserving for SPECT or PET modality, but such modality-preserving embodiments are also conceivable that relate to other modalities; furthermore, inthe modality-preserving case, in line with the application of a modality-preserving reconstruction module, the forward projection module applied during the training method will also be modality-preserving).Also in accordance with the above, it holds true for the embodiment illustrated in Figs. 1 -5 that a SPECT or PET modality is applied in the system, and that in addition to providing medical visualising data packages (e.g., medical images) as an input to the reconstruction module and volumetric visualising data packages (e.g., volumetric images) as an input to the forward projection module, related CT anatomy information or MRI anatomy information (in other words, a CT map, or alternatively an MRI map could also be applied) is also provided. Taking into account CT or MRI anatomy information for the machine learning reconstruction module and forward projection module can be implemented for example by applying multi-channel-type neural networks. Accordingly, the neural network receives a larger amount of initial data (i.e. from several data sources), which - and the training method - will select the relevant data from these.In the following, a system of neural networks preferably applied in the modules and their training process will be described. Special aspects related to the embodiments applying quantum neural networks will be addressed in the description of the individual stages.Stage 1 relates to generating the training database. The training database (of images or visualising data packages) can be generated in various ways; in the following, a preferred embodiment of this, based on Monte Carlo modelling, will be described. In stage 1 , various methods having basis other than machine learning (in other words: free from machine learning, non-machine learning-based, nonmachine learning, (analytic) calculation-based, do not comprise machine learning; briefly: non-machine or conventional, classical) can be used for generating the training database; the corresponding module is termed a different-from-machine learning-based reconstruction module (the other adjectives introduced above can also be used in the name of the module). It is emphasised that in this sense it is determined: the module must be suitable for reconstruction and it is required that, of the groups of machine learning-based and different-from-machine learning-based modules it belongs to the latter. In addition to Monte Carlo modelling, other different-from-machine learning-based methods (i.e., methods based on solving / integrating equations) are also suitable for generating the training database, such as for example raytracing. Of such integration techniques, the Monte Carlo method can be considered the preferred one with more than 2D.The aim of this training stage is to generate an image database that is appropriate for training the neural networks. For example in case of SPECT / PET, which modalities are illustrated in Figs. 1 -5 (and there is modality-preserving, i.e. no modality switching at the reconstruction), the Monte Carlo-based reconstruction (the module designated with the label “MC rec.” - i.e., “Monte Carlo reconstruction” - in Fig. 1 , that is, a reconstruction module 100) is applied only at this stage (phase), so it does not slow down the actual training process, because the generating of the training database is a task that can be separated, i.e., the database may be made available in advance. The process of generating the training database is illustrated in Fig. 1 . By applying stage 1 as illustrated here it can be made possible that the measurement time and the X-ray dose or the dose resulting from by the injected isotope the patient is exposed to can be significantly reduced.In case of constructing the training database according to Fig. 1 , we start from PHmedical images of a base set 110 - which in the embodiment illustrated in Fig. 1 are projection images (SPECT) or LOR-s (line-of-response, in case of PET measurement) having good statistics - i.e., from measured information. These comprise measurements performed according to the given examination protocol. The situation is the same in case of CT and MRI.From the medical images of the base set 110 (in an alternative term, “base training set”) a splitter module 105 (a so-called “splitter operator” with a splitter index j) generates measurements with shorter measurement time / lower dose, generating j measurements from each image (these will become PLj medical images of a first auxiliary pre-training set 115, the index j running through the medical images belonging to a given PHmedical image (base image); of course there is a plurality of PHmedical images as they make up the base set 110). In case of SPECT and PET, this can be performed based on a draw following a binomial distribution; the PLj measurements characterizable by lower signal-to-noise ratio and a relatively lowimpact number, i.e. , the medical images of the first auxiliary pre-training set 115 are generated this way.The preferably high-accuracy, conventional (i.e. not implemented by a neural network) reconstruction module 100 operator is designated with “MC rec.”, which generates VHPvolumetric images of a second auxiliary pre-training set 120. These can be 3D, i.e., spatial data, or 4D data that also have time dimension in addition to the spatial dimensions; in the case of the latter the time dimension is also included in the input data (medical visualising data). The optional application of the time dimension extends through all stages, so the training data - i.e., also the training data of stage 4 - are also in compliance with this. Of course, the time dimension is present in the training process if the reconstruction applying the trained system is to be applied for time-dependent data. A parameter p referring to a given reconstruction setting (i.e., a given combination of the parameter values of the reconstruction parameters) is applied. In case of an iterative reconstruction solution (e.g. EM or OSEM, that is, “expectation maximisation” or “ordered subset EM”), the reconstruction parameter 102 p is meant to include the number of iterations and any other parameters of the model, including the parameters of the internal filters. Reconstruction is performed applying the reconstruction module 100 for different combinations of the reconstruction parameters, thereby obtaining reconstructions carried out in different ways. Thus, several different reconstructions (all considered appropriate) will be available during the training in stage 2, wherein it can be rewarded if the reconstruction module 200 obtains a result that is similar to one of them. The medical field (application) is not yet constrained here; as the preferred aspect of the reconstruction is not known yet, all aspects are rewarded. The advantages of this, see at stage 4.Preferably CT anatomy information 125 is also utilised for generating, which provides additional anatomy information for medical images (PHmeasurements) having SPECT or PET modality. This CT anatomy information 125 can be also used for the training data in stages 2 and 3 (see Figs. 2 and 3), while for stages 4 and 5 it is necessary to generate new CT anatomy information.If the modality of the medical image to be reconstructed is CT or MRI, then obviously additional CT or MRI anatomy information is not fed either to the reconstructionmodule 100 in stage 1 or to the reconstruction module 200 and forward projection module 300 in stages 2-5. This constitutes the difference in the case of CT or MRI modality, otherwise everything remains the same in the illustrated embodiments.Considering CT (or MRI, for the sake of simplicity hereinafter only CT is mentioned) anatomy information, we have the following options (hereinafter, modalities other than CT (or MRI) are observed because in the CT and MRI modalities CT or MRI anatomy information is included):- modality different from CT, but CT anatomy information is not applied at all in any stage;- it can be done, if stage 1 is applied, CT anatomy information is also applied therein, but in the other stages CT anatomy information is not used (because, although it could be carried over from stage 1 to stages 2-3, for stages 4 and 5 the appropriate CT anatomy information would somehow need to be obtained).- it can be done, that CT anatomy information is used both in stage 1 and in stages 2-5: expediently the same information is used throughout stages 1 -3 (because it belongs to the training data), in this case CT anatomy information must also be obtained for stages 4-5.Accordingly, in the course of the preparation, reconstructions are performed by reconstructing the PHmeasurements (medical images of the base set 110) that were taken with the normal protocol (and thus have relatively high signal-to-noise ratio) with different reconstruction parameter sets p. Modifying the parameter p is meant to include choosing a different number of iterations or different internal filter parameters. The model is always as accurate as possible, so, e.g. in the case of Monte Carlo modelling sufficiently high photon numbers are selected for forward and backward projections and for modelling scattering. The reconstruction module 100 (the “MC rec” operator) applied here is not optimised for speed.In the embodiment described above, that is, applying the numbering introduced above, in the so-called stage 1 (which precedes the pre-training of the reconstruction module and the forward projection module in accordance with the objective of stage 1 , i.e. generating the training database)- in the first auxiliary pre-training set of the medical visualising data packages (see e.g. medical images of the first auxiliary pre-training set 115 in Fig. 1 as medicaldata packages of an exemplary first auxiliary pre-training set), pre-training auxiliary medical visualising data package groups comprising medical visualising data packages with different noise levels (such for example images with different noise levels are in an image group as a data package group) each generated from a base set of medical visualising data packages (see e.g. medical images of a base set 110 in Fig. 1 as medical visualising data packages of an exemplary base set) with a splitter module are applied, and in the second pre-training step, as the visualisation information corresponding to respective given medical visualising data packages of the first auxiliary pre-training set of the medical visualising data packages, from the base set of the medical visualising data packages, the medical visualising data package corresponding to each of the medical visualising data packages of the first auxiliary pre-training set of is applied, and- the second auxiliary pre-training set of the volumetric visualising data packages corresponding to the first auxiliary pre-training set of the medical visualising data packages is generated by processing each of the medical visualising data packages of the base set of the medical visualising data packages by means of a different-from -machine learning-based reconstruction module (see e.g. a reconstruction module 100 which is based on Monte Carlo modelling).In this case, by referring to the second auxiliary pre-training set of the volumetric visualising data packages corresponding to the first auxiliary pre-training set of the medical visualising data packages it is meant that the volumetric visualising data packages of a second auxiliary pre-training set 120, which have been generated from the base set 110, correspond to the medical visualising data packages of the first auxiliary pre-training set 115 which have also been generated from the base set 110. It can be understood that, applying the splitter index j and the parameter p visualising data packages originating from the same medical visualising data package of the base set 110 correspond to each other. These mutual correspondences are therefore determined.Preferably, furthermore (that is, as an option with respect to the above)- for each of the medical visualising data packages of the base set of the medical visualising data packages, a plurality of volumetric visualising data packages ofa second auxiliary pre-training set (see in Fig. 1 e.g. volumetric images of a second auxiliary pre-training set 120 as an exemplary plurality of such volumetric visualising data packages) are generated by processing those with a plurality of parameter values being different from each other of reconstruction parameters (see a reconstruction parameter 102 in Fig. 1 ) corresponding to the different- from-machine learning-based reconstruction module (see a reconstruction module 100 in Fig. 1 ; these operations are carried out in stage 1 , the results being used in stage 2 below), and- the first pre-training loss function (this can be metric: a functional measuring the distance between the two images is established, or in many cases it can also be based on difference) is composed based on a plurality of volumetric visualising data packages of the second auxiliary pre-training set generated by processing with a plurality of parameter values different from each other of the reconstruction parameter and on first pre-training volumetric visualising data packages (see e.g. the volumetric images 210 in Fig. 2 as exemplary first pretraining volumetric visualising data packages) generated by processing the medical visualising data packages of the first auxiliary pre-training set by means of the reconstruction module (see the training scheme illustrated in Fig. 2 for stage 2, wherein the results are processed according both indices j and p in the loss function).Thanks to the different reconstructions generated in this way, stage 4 can be even more effective, because it becomes possible to reduce a larger latent space (that is richer in details) in that stage. Therefore, the latent space can be extended using the splitter module, and by applying different reconstruction parameter values (these extensions of the latent space may be performed independently of each other).If a quantum representation is applied in the subsequent stages, then the reconstructed VHPvolumetric images (in other words: volumes) are transcoded into a qubit representation. That is, in such cases the VHPvolumetric image is an appropriate ket state |VHP>, i.e., a vector representing a quantum state, and in a similar fashion a state |p> can be produced for the p medical image (here as well, images of all types can be generalised to visualising data packages, i.e., the quantum representation can be applied generally for visualising data packages).In stage 2, pre-training of the reconstruction module is performed, i.e., pre-training of preferably an end-to-end reconstruction autoencoder, or of an equivalent neural network topology, e.g. a visual transformer (see also below) or of generative networks.In this stage the reconstruction module 200 is pre-trained based on the available PLj measurements (medical images of a first auxiliary pre-training set 115), and based on reconstructed VHPimages (volumetric images of a second auxiliary pre-training set 120) paired with them, that is, for example a neural network (designated as RECNN, reconstruction neural network, in Fig. 2) having an appropriate autoencoder architecture (e.g. Il-net) is trained (pre-trained) such that it can fulfil the reconstruction task. In the reconstruction module 200 there can be seen an encoder 202a and a decoder 202b. All possible reconstruction settings denoted by p will be encoded in the established latent space, i.e., by means of the training process the domain and the range of the representation space established downstream of the encoder is defined. The encoder selects the minimal but important feature. In other words, it finds the “essence”, i.e., the minimal representation of the required image features. The decoder, on the other hand, is a generative neural network that generates from an element of the representation the corresponding reconstructed image.A flow diagram of the process of pre-training is illustrated in Fig. 2. Adjusting the parameters of the neural network, that is, a so-called first pre-training to be applied in stage 2, is indicated by an arrow corresponding to operational step S220; training is carried out (double arrow at the start of operational step S220) based on a loss function calculated based on reconstructed vfilteredj volumetric images 210 generated by the reconstruction module 200 and on VHPvolumetric images of a second auxiliary pre-training set 120. The loss (i.e., this is the first pre-training loss function) can be typically calculated using either L1 or L2 metric, but other loss functions may also be applied. In this case, training is performed in a supervised manner.In the multichannel autoencoder neural network exemplary applied in stage 2, in accordance with the application of multiple channels, a supervised co-learning (collaborative learning) strategy is implemented (Zhaoheng Xie et al.: Anatomically aided PET image reconstruction using deep neural networks, Medical Physics 2021Sep; 48(9): 5244-5258). As a result of this, an encoded latent reconstruction space (corresponding to healthy and sick (having disease)) representing a large population is generated. It may be preferable if this space is sufficiently dense in order that no false structures are produced during decoding, see for example the variational autoencoder-type neural nets (Xianxu Hou et al: Deep Feature Consistent Variational Autoencoder, 2017 IEEE Winter Conference on Applications of Computer Vision (WACV)) the application of which constitutes a further preferred embodiment in addition to multichannel autoencoders. A further preferred embodiment is constituted e.g. by visual transformers (A. Dosovitskiy et al: 'An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale". arXiv:2010.11929v2, 2021 , see also below).Applying a full quantum representation, according to the above, i.e. , if a quantum neural network is applied in the reconstruction module 200, the REC-NN is an operator acting on an image state converted into the ket state |PLj> that produces the state |Vfiltered>. The loss function (error function, Qloss) that is to be minimised during the training process is a complex distance calculated for the probability coefficients of the ket bases of the image space, i.e., a distance calculated on the Bloch hypersphere applied for image representation, which is to be calculated (summated) for all parameters j and p, that is:Thus, the above discussion should be taken as applying a quantum neural network in the full quantum representation: in hybrid quantum representations other training methods may be utilised, wherein e.g. classical comparison is applied in the loss. In this alternative solution, we would also like to realise the state |vfiltered> as a real image. In such a case, the state |vfiltered> must be projected to a base vector <xyzo| of the image space, that is, the volumetric image is obtained as an inner product (scalar product): vfiltered= <xyzo|Vfiltered>. In this case, the loss is calculated according to the loss calculation used for conventional neural networks (for example, according to L2 metric), i.e., comparing the Vfilteredand VHPvolumetric images and summating them for all parameter values j and p.In stage 3, pre-training of the forward projection module, that is, preferably the pretraining of a forward projecting neural network (FP-NN) is performed. Subsequent to stage 2, therefore, in the illustrated embodiment a forward projection module 300 (implemented by a neural network in the illustrated embodiment) is pre-trained, which is adapted for projecting a reconstructed Vfilteredj volumetric image 310 - generated by a reconstruction module 200 already trained in stage 2 - to the space of measured data representations (in the illustrated embodiment: in the case of SPECT, projection images, in the case of PET, LOR-s). The pre-training process is illustrated in Fig. 3.Training is performed in a supervised manner also in this case (like in stages 4 and 5), i.e., based on PLj (medical image of the first auxiliary pre-training set 115), -> Vfilteredj (volumetric image 310) Pfilteredj (medical image 320) pairable with data packages / medical images PH(medical image of the base set 110); the training is aimed at reducing the difference between PHand pffteredj |_|ere as we|| ^g |ossfunction can be calculated based on the L1 or L2 metrics (or can be calculated otherwise, this is the second pre-training loss function), with which the parameters of the FP-NN network are determined by backpropagation or other algorithm (see the arrow indicating operational step S330 in Fig. 3).It is important to emphasise here that according to this arrangement the module assembly (entity) constituted by the reconstruction module 200 and the forward projection module 300 is pre-trained in stages 2-3 such that the reconstruction module 200 is pre-trained first, and it is then applied in its pre-trained state for pretraining the forward projection module 300 (the training process requires an already functioning reconstruction module 200).As illustrated also in Fig. 2, in the training of the reconstruction module 200 in stage 2 two running indices meet at determining the loss function, i.e. the index j (corresponding to the noise level) and the index p (corresponding to the parameters applied for the evaluation). In order to determine the loss function, all values of both indices are applied (i.e. summation is performed for j and p), that is, the training of the reconstruction module 200 implemented by a neural network according to Fig. 2 is performed for each j and p. Thanks to that, the reconstruction module 200 istrained in general, i.e. , thanks to a wide training database by the help of that the interpretation range and the value set range is covered.In stage 3 training, summation is done for the index j (i.e., in Fig. 3, a comparison is preferably performed with PHrather than with PLj; training of the forward projection module 300 implemented by a neural network according to Fig. 3 is performed for each j), such that the system tolerates different noise levels. If a fixed j, that is, a given image statistic (noise level) was fixed, the interpretation range of the forward projection module would be restricted.In the invention, thanks to the applied process of the training, the interpretation range becomes restricted only in stage 4 and in the optional stage 5, the result becomes accurate then.In stage 3 the parameter p has no role (as there is no place where VHPcould be inserted), i.e. the invariance from p has already been encoded into the reconstruction module. Comparison is performed with PHin stage 3 because we would like to make backward projection statistics-independent.Applying quantum representations (i.e. quantum neural networks), as an embodiment, the FP-NN neural network (forward projection module 300) built from quantum neurons or parametrisable quantum gates is an operator that preferably acts on the quantum state |Vfilteredj>, with the ket state |pfi|teredj> being its output. Similar to the Qloss above, this is compared with the state |PH> generated from the PHimage.In stage 4, medical field-specific training is performed, further training the reconstruction module and the forward projection module as set forth above. In the illustrated embodiment the details of such a training will be described below. At the end of stage 3, the image obtained through reconstruction (in general: visualising data package) is blurred in many cases - because it is not an objective at this stage to generate an excellent-quality reconstruction -, that is, it is not really appropriate for any given medical field (application). However, this is a valuable state that is in accordance with our objectives: while preserving the pre-trained state (thus, mightas well maintaining the possibility of returning to it), it can be specifically restricted to various medical fields (applications) in stage 4.In the present stage 4 the reconstruction module 200 and forward projection module 300 (REC-NN and FP-NN networks) pre-trained in the previous two stages (stages 2 and 3) are trained further. The objective of this stage is the incorporation of medical field-specific information such that the system of the two neural networks can find the reconstruction setting(s) that is optimal for the given medical field. Essentially the range of latent representations becomes restricted or specialised in this step.The medical field-specific artificial investigation area models to be applied in stage 4 are aimed at preparing for stage 5 as well as possible such that in stage 5 only fine-tuning will be required. Accordingly, when preparing the training data based on such established modelling, we seek to have various sizes among the models (body sizes, bigger or more slender body shape, etc.), different anatomical and bodily characteristics, the models (phantom sets) are ideally made with great variety from the aspect of activity (its distribution, the amount of contrast agent).In accordance with this embodiment of the invention a quantitative evaluation method is preferably defined for the further training of the neural networks, that is, one or more biomarkers to be calculated (i.e. , we focus on specific abnormalities or alterations related to the given medical field, see below). This will be illustrated below by an example; we also refer to the detailed disclosure included below wherein the loss function expediently applicable in stage 4, i.e., the details of the evaluation method are presented.Describing a particular non-limiting example, e.g., the targeted medical field is heart examination by SPECT (for which field-specific training is to be performed in stage 4), and the parameter to be measured by the radiologist is the heart muscle perfusion of the left ventricle, through which the status of the cardiac muscle (i.e., presence of a potential ischaemic heart disease) is wanted to be characterised by the activity reduction (decrease, lack, absence) rate in the cardiac wall; in such a case it is necessary to define for example a method of determining the activity uptake and distribution of the healthy uptake, i.e. healthy heart muscles, and measure against it the amount, contrast and size of the reduction. In such a case,for example the mathematical phantoms named XCAT (the basis of the name: extended Cardiac Torso) can be utilised (see Segars WP, Sturgeon G, Mendonca S, Grimes J, Tsui BM. 4D XCAT phantom for multimodality imaging research. Med Phys. 2010 Sep;37(9):4902-15; hereinafter: XCAT study), wherein the activity distribution of the organs (and temporal change thereof) and specific reduction in the cardiac wall can be parametrised (for example, various reductions can be inserted into the cardiac wall, modelling a variety of artery (vessel) diseases). Then we can specify a measuring number adequately representing it, for example the contrast value of various segments of the left ventricle.In Fig. 4, VXCATk denotes the 3D (or 4D, if time-dependent training data are to be applied) volumetric image, taken from the medical field-specific training set 410, of a mathematical phantom according to the present example that is very similar to the human morphology, and the index k may denote different organ sizes and patient models and activity reductions with different measure, size, and location inserted in the cardiac wall. In stage 4 this constitutes the training data.Together with the activity volume (i.e., volumetric image of the medical field-specific training set 410), CT anatomy information 425 (CT map, material map) denoted with CTXCAT similar to a CT is also recorded, which also constitutes input information of the preferably multichannel-type neural networks REC-NN and FP-NN, i.e., the reconstruction module 200 and the forward projection module 300. It is noted that if the base modality is CT (for example, Fig. 4 corresponds to SPECT or PET base modality, and in line with the modality-preserving, VXCATk images of the medical fieldspecific training set 410 also have this modality), then the XCAT or other similar phantom will only produce the CT mathematical phantom, and the VXCATvolume in Fig. 4 will be a mathematical CT (obviously without an additional CT map).As shown also in Fig. 4, the reconstruction module 200 and the forward projection module 300 (REC-NN and FP-NN) undergo further training together (arrows corresponding to operational steps S440 and S450) such that the loss between the ^ / filtered volumetric image 430 (medical field-specific training loss-check volumetric image 430) and the VXCATk volumetric image of the medical field-specific training set 410 is minimised (see the double arrow between these corresponding to the medical field-specific loss).In this case, training is performed in stage 4 by starting with feeding a VXCATk volumetric image of the medical field-specific training set 410 to the forward projection module 300, which processes it and generates a forward-projected Pfilteredmedical image 420, which passed to the reconstruction module 200, it generates the reconstructed vfilteredvolumetric image 430. This can be compared with the VXCATkvolumetric image of the model-based medical field-specific training set 410, i.e., it can be checked how the elements of the system (in the following order: forward projection, reconstruction) have affected it. Having performed the training according to stage 4 - as stage 5 is only optionally included - the output of the system may be constituted by the reconstructed vfilteredvolumetric image 430, as the volume reconstructed by the reconstruction module 200 (volumetric image; see Fig. 4). It is noted that the index k used in VXCATk - as it denotes the k-th element of a set - could also be taken over to the CT anatomy information blocks 425, the medical image 420, and the volumetric image 430 because they also correspond to the k-th element of the set. However, in Fig. 4 there is no such index comparison as was shown in Figs. 1 -3, so here the additional indices k are not shown (the index k could also be shown in Fig. 10 for these).Preferably, an XCAT phantom is also applied during training, in an example both as an artificial investigation area model and during the preferred training verification (the latter may be applicable in this stage for verifying that the training has been sufficient, i.e., for stopping the training process). During verification the phantom is expediently generated applying other parameters so that to be an image that has not yet been “seen” by the network being trained.The loss applied in stage 4 compares medical field-specific biomarkers (i.e., it compares an abnormality or alteration designated according to the given medical field with its healthy counterpart), i.e., it is typically not based on voxel-voxel distance (for detailed description see below).A consequence of the joint training (common training) of the two neural networks (REC-NN and FP-NN) is that the REC-NN operator starts to generate reconstructions that are optimal from the aspects of the given medical field; that is, the latent space containing the reconstruction structures evolves during the trainingsuch that a reconstruction parameter setting that is optimal for the given medical field is also established.Performing the reconstruction in stage 1 applying preferably various combinations of parameter values provides the following advantage for stage 4. Because in stage 1 reconstruction is performed for various combinations of parameter values, and, for each of them, those outputs of the machine learning-based reconstruction module 200 that are sufficiently close to them are rewarded in stage 2, during medical field-specific further training, when we already have the specific aspects and the medical field for which the reconstruction should be optimised, the reconstruction module 200 can be restricted from a larger set of trained ones (i.e. , the latent space can be reduced this way), i.e., an appropriate combination of parameter values can “select itself”.A more detailed description of stage 4 is given below. For an appropriate performing of stage 4 appropriate training data are required. The training data can be for example generated in accordance with the XCAT study; however, a generalisation of that is given and other possibilities for generating training data.In relation to that it is also important that the XCAT study (for example, but also see below for other options) provides training data input to the invention, i.e., to the training method constituting an essence of the invention. In other words, the outputs generated with the help of the XCAT study can be for example applied for training - i.e., exclusively as an input - in the system according to the invention. The XCAT study does not contain several important aspects (training stages, machine learning modules and their “cooperation”, etc.) of the invention.With the help of the XCAT framework, the data required for training, that is for example the healthy and abnormal portions of the dataset can be obtained either directly or with minor additions (completions).The XCAT approach generates 3D (or 4D) mathematical phantoms - i.e., for example a 3D voxel array - based on a parameter file. The training dataset is generated by modifying the parameters. Parametrisation allows us to directly insert an abnormal portion therein applying the XCAT approach, for example byappropriately tuning activity at specific locations (i.e. for example, allows the addition of different activity values into the putamen brain regions in the left and right hemispheres, which allows the modelling of dopamine transporter deficiency, i.e., Parkinson’s disease). In another example, for example, it will consider the left ventricle to have homogeneous activity distribution or to have homogenous physiological functions, that is, it will generate a healthy heart. In relation to that an reduction can be “inserted” to the cardiac wall with the XCAT package, i.e., coronary disease can be modelled.Certain abnormal alternations, such as bone lesions and metastases cannot be directly modelled with the XCAT framework. In such cases, the results provided by the XCAT framework can be complemented, inserting for example various accumulations to the bones. That is, a specific accumulation image is generated by XCAT, and, depending on the medical field, either this result is used as is or this result allows us to also generate the sick image (image with disease; i.e., the abnormal portion of the volume).The dataset obtained according to the above may also be augmented, i.e., the number of volumetric images of the available training dataset may be increased by applying shifting, rotation, shrinking and zooming to the images (see the preparation for stage 5 above).Thus, there are known approaches for producing such training data generated based on artificial investigation area models. In addition to the XCAT study, there are other available possibilities, such as the so-called Zubal brain phantom (see for example Zubal, I.G., Harrell, C.R, Smith, E.O, Rattner, Z., Gindi, G. and Hoffer, P.B., Computerized three-dimensional segmented human anatomy, Medical Physics, 21 (2):299-302, 1994.). All other anthropomorphic mathematical phantoms can also be suitable for the purpose. Also, thinking in other directions, artificial investigation area models may also be generated for example by segmenting high-resolution measured and reconstructed images, or applying affine transformations to these images.Of the known approaches, therefore, very high-quality training data can be generated based on the XCAT study, but depending on the medical field other -even less sophisticated - approaches may also be applied for generating the training data (e.g. the XCAT study also contains references to the NCAT approach, which is an earlier version of XCAT). XCAT is more detailed than NCAT, and more closely resembles a human body. However, several extensions have also been made to it applying GAN neural networks, so it can also be a feasible choice for generating the artificial investigation area model (Chang Y et al.: A generative adversarial network (GAN)-based technique for synthesizing realistic respiratory motion in the extended cardiac-torso (XCAT) phantoms. Phys Med Biol. 2021 May 31 ;66(11 )). However, their application in the training method according to the invention, that is, for training the system, does not appear in the known approaches. That is, the approaches adapted for generating training data are applied simply as tools for generating a training database.The XCAT study is able to model all the organs (e.g. heart, brain) that are important from the aspects of the invention. If other aspects are to be included into the model, the approach set forth in the XCAT study can be supplemented (for example, temporal changes of metabolism between organs); it also does not take into consideration intra-population variations, but it can be developed further also in this regard. In general, it can be supplemented as necessary depending on the given medical field, but the mode of implementation of this does not form a part of the invention.The organs that are important from the aspects of the invention are bone, kidneys, brain, the heart, as well as tumours occurring anywhere in the body - for example bone metastases, soft tissue metastases, etc. Moreover, preclinical examinations of small animals (mouse, rat, rabbit, monkey, pig, etc.) are also important as physiological and translational models for the field of human examinations.At the most general level, in stage 4 an artificial investigation area model (alternatively, the adjective “artificial” can be replaced by the adjective “synthetic” in the term) is applied as training data, which is made of an examined portion of the body, i.e. , it corresponds thereto. As an alternative to the term “artificial investigation area model” (artificial model) the term “phantom model” can also be used, which is widely used in the literature; in the present application it is used in an equivalent sense to the term “artificial investigation area model”. The artificial investigation areamodel - as it is set forth in the XCAT study - can be voxelised, mathematical (i.e. , defined by equations and with the help of geometrical shapes), or a combination of the two. It is noted that the voxels are applied in a given representation. However, other complete base function systems can also be applied for describing the activity and morphology of the body. Such are e.g. the so-called “blobs”, but any 3D wavelet is also applicable (A Yendiki and J A Fessler: A comparison of rotation- and blobbased system models for 3D SPECT with depth-dependent detector response, 2004, Phys. Med. BioL 49, 2157). The phantom described in the XCAT study is a voxelised mathematical phantom, i.e., a parametrisable voxelised model. As long as we apply a pixel / voxel-based image representation, also obtaining the generated volume in the same representation is preferred. In the quantum-based case, the inner representation uses qubits, but it is also converted to a conventional representation (see also at the description of the quantum-based aspects).In stage 4, the training data are constituted by volumetric images generated based on a medical field-specific artificial investigation area model (see in Fig. 4 the VXCATk volumetric images of the medical field-specific training set 410). If CT anatomy information is applied for the reconstruction module 200 and the forward projection module 300 (see the CTXCATCT anatomy information 425 in Fig. 4), then this CT anatomy information corresponding to each of the reconstructed training volumetric images must also be generated.For defining the loss function that is to be expediently applied in stage 4, the contrast-to-noise ratio (CNR) is introduced:in which CNR formula the function “mean” and “std” stand for the arithmetic mean of the voxels falling into the given VOI (“volume of interest”, i.e. volume that is of interest, or, shortly interesting volume) and the standard deviation of the voxel values belonging to the given VOI, respectively. In the formula the healthy and abnormal VOI-s relate to different volumes (see also below). In a general case, i.e., applying visualising data packages (that is, when there can be time-dependence), time dependence may constitute an additional parameter, for which expedientlysome type of averaging is applied (for example simple averaging (arithmetic mean) or sliding-window averaging), so time dependence can also be taken into consideration in the formulas.The approach set forth in the XCAT study can be considered a “volume generator” wherein the VOI-s can be determined. For example, in the approach according to the XCAT study such settings can be applied that only the cardiac wall of the left ventricle or the appropriate regions of the brain are generated (see the related definitions of the VOI-s below), nothing else. The VOI includes all such regions that are of interest from the aspect of evaluation. Also, in case the temporal variation of the mean activity can be experienced in the VOI-s (as averaging is performed for the entire VOI) is also of interest (4D application), then time dependence is also taken into consideration in the loss (i.e., averaging is carried out for all time instances taken into consideration whereby noise reduction can be achieved).Based on the contrast-to-noise ratio (CNR), preferably the following loss function can be constructed (i.e., based on the indices of the reconstructed training volumetric images, that is, summated for each image pair; this is illustrated by an example of cardiac wall-reduction, however, we maintain the generality of the above CNR formula, the application of which in the loss helps reconstruction characterizable with high signal-to-noise ratio):In line with the application of the CNR formula set forth above in the loss, in the present example the cardiac wall gives the healthy term, and the reduction gives the abnormal term (for the example of cardiac wall-reduction corresponding to this formula see also below the identification of the healthy and abnormal VOI-s). It is important to note that the two “mean”-s in the formula relate to different VOI-s, the first relates to the cardiac wall where there is no reduction (i.e., in general, to the healthy region), and the second relates to the space part of the place of the reduction(i.e. , in general, to the region affected by the abnormal alteration). Each volumetric image contains both the abnormal and the healthy regions, although at different parts of the image. Therefore, the loss contribution corresponding to each VXCATvalue can be calculated starting from it. The volumetric image (in this case, XCAT phantom) is therefore the same in the healthy and abnormal case, and the VOI is set at a different location (the location of the healthy and abnormal regions is known as they have been generated by us).Thus, referring also to the CNR formula above, by rearranging this formula it can be seen that in the loss value obtained in such a manner the XCAT and the filtered components originating from the healthy and the abnormal regions can be compared (by rearranging the formula, such differences can also be obtained in the formula), i.e., by applying the loss formula both are measured (preferably, in this way, simultaneously), that is, the quality of reconstructing the healthy and abnormal parts in Vfilteredcan be got back. The locations of the VOI-s are chosen based on the XCAT training data generated by us (thus we know where the healthy and abnormal regions are) and on the given medical field (application). In case of healthy training data, the abnormal XCAT term is eliminated from the definition above: in such a case no abnormal portion is marked in the training data, so the corresponding term of the formula is zero (the VOI is placed where an abnormality would occur in the case of abnormality-containing training data, because the location and the type of potentially occurring abnormalities are typically specific in an artificial investigation area model, depending of course also on the modality). Therefore, for such training data the loss function measures how well the healthy data are reconstructed; typically, there will be an abnormal term in the VOI in the data to be compared (Vfiitered jnpjg 4)a|sojn suc cases because during training the system applies imperfect imaging.It is noted that the accumulation may show up as an addition and may be an accumulation deviating from normal (lesion), but reduction may also be considered as an accumulation deviating from normal: for example, a given space part is fainter than in a healthy state, that is, it is a type of “negative accumulation”.Also, in accordance also with Fig. 4, in the loss function a comparison between the reconstructed VXCATand Vfilteredvolumetric images is to be made. Therefore, theapplication of the above-included CNR formula appears in the formula of the loss function: the CNR is calculated for both reconstructed VXCATand Vfilteredvolumetric images, the loss being obtained as the difference of the two.The formula includes “mean” because each pixel of the VOI (put on the cardiac wall) must be taken into consideration. In turn, there is “std” because it is also important that a homogeneous organ part (e.g. the cardiac wall) remains homogeneous. That is, in the example cited above: we would not only like to have a contrasty view of the activity reduction in the heart (i.e., generally, of the abnormal alteration), but would also like to ensure homogeneity. But, since the “std” of the original phantom - i.e. the one generated with XCAT - calculated for the cardiac wall is low, of the XCAT and the filtered values preferably the filtered one is included in the numerator.In the formula, VOI(volume, mask) is a function that acts on the examined volume, and only retains the pixels corresponding to the given mask and discards the remaining pixels; the mean and std run through only those pixels that the VOI function has retained. Sticking with the example related to the heart muscle, the VOI put on the heart muscle includes those voxels (i.e., the function corresponding to the VOI retains these voxels) that include the cardiac wall of the left ventricle (this is considered the healthy VOI). In turn, the VOI taken onto the reduction (abnormal VOI) retains those voxels that include the reduction taken onto the cardiac wall of the left ventricle (that is, a healthy and an abnormal region of the cardiac wall is included).In this embodiment, therefore, preferably in the medical field-specific training step, by composing the model reproduction loss function based on the volumetric visualising data packages of the medical field-specific training set generated based on the medical field-specific artificial investigation area model (see e.g. in Fig. 4 volumetric images 410 of the medical field-specific training set as exemplary volumetric visualising data packages of a medical field-specific training set generated based on the medical field-specific artificial investigation area model) and on the medical field-specific training loss-check volumetric visualising data packages (see e.g. medical field-specific training loss-check volumetric images 430 in Fig. 4 as exemplary medical field-specific training loss-check volumetric data packages)- first deviation data and second deviation data (in the above embodiment these are the first and second contrast-to-noise ratios; the deviation data is composed of data elements - e.g. numeric values contained in each voxel - originating from the visualising data packages, i.e., it provides deviation for each voxel) characterizing deviation from healthy (cf. with descriptions about biomarkers) are calculated for the volumetric visualising data packages of the medical fieldspecific training set generated based on the medical field-specific artificial investigation area model, and for the medical field-specific training loss-check volumetric visualising data packages, respectively, and- the model reproduction loss function is determined as a deviation (difference) between the first deviation data and the second deviation data (deviation refers to a positive numeric value derived from the difference; the determination can be done by taking the absolute value or applying an even power, i.e., the model reproduction loss function is given by taking the absolute value of the difference of the first contrast-to-noise ratio and the second contrast-to-noise ratio).The CNR-based loss included above is suitable for all modalities and medical fields, but the above-described example relates to the calculation of a loss function specific to perfusion heart examination. In the case of other examinations, the loss value is to be calculated for the organ in question or the specific biomarker which can be evaluated by the doctor, and, in stage 4, further training of the neural networks is performed accordingly. However, particularly preferably, the embodiments based on the first and second deviation data provide a quantitative evaluation method.For example, in case a DatScan examination is carried out for determining by scanning the striatum of the brain if a patient has Parkinson’s disease, the VOI will be applied as follows: the abnormal VOI will be set to the putamen / caudate region, while the background, i.e., the healthy VOI will be set to the white matter of the brain as the latter is presumed to be a healthy region. However, other biomarkers and index numbers can also be applied for the VOI, i.e., in the medical field-specific training step the specific regions, the abnormal deviations occurring therein that are to be put into the focus of the system can be selected by selecting the VOI, and, with the help of the loss, the quality of reconstructing the abnormalities by the system can also be controlled, while the reconstruction ability of the system can beimproved applying the stage 4 training method detailed above. Accordingly, by “taking into consideration a biomarker” it is meant that what type of abnormality or deviation the training is prepared for (see the formula above).In connection with the approach of the XCAT study or other selected “volume generator” with similar functionality it is noted - illustrated with an example related to the cardiac wall - that these approaches include a module that is able to take a reduction to the heart muscle of the left ventricle (or, in general terms, to apply a change between the abnormal and healthy state, i.e., to apply an abnormal deviation). In such a manner, the deviation from the healthy state can be appropriately represented and controlled in the generated volumes.It is important to note in relation to the different training stages, i.e., for example to switching from stage 3 to stage 4, that the reconstruction module and the forward projection module (especially the latter, since the pre-training of the former has been stopped earlier) actually “realise” that they are in stage 4 due to the modified expectations, i.e., due to being penalised and rewarded for different things than before (since the loss is calculated differently), and the modules will learn in accordance with this. Thereby, switching from stage 3 to stage 4 itself “brings out” medical field-specificity from the reconstruction and forward projection modules, as at stage 4 we begin to train the modules to focus on a given medical field (an organ, a type of abnormal deviation, etc. specific to the medical field), that is, what counts in the stage 4 is how they perform in the given medical field.Prior to stage 4, the system was trained with the j and p indices being “left open”. Therefore, it has been trained in a generic manner, in other words, in a way that it is more suitable to undergo medical field-specific training. Applying training according to stage 4, the space where the system searches for the solution (the reconstruction) is constrained. In other words, by applying further training of the system according to stage 4, its output is made more accurate relative to the output that can be obtained by the previous generic training (which is given by the reconstruction module otherwise).The phantom, the applied model is always related to the given medical field, that is, in both the healthy and abnormal case a given body region or a given organ can beseen therein. More particularly, multiple types of phantoms (models) are generated (for example for different body sizes and organ sizes, etc.), and the parameter k shown in Fig. 4 runs through the various generated phantoms.In case of applying a quantum representation in stage 4, i.e. , in case of applying quantum neural networks in a preferred embodiment, because it is not feasible to calculate the medical field-specific loss with simple quantum operators, for example the loss characterised with the auxiliary function f(a,b) (this is the model reproduction loss function in quantum representation, its formula is identical to that of the function applied in the classical case; on how to make the required substitutions see below), of which we search the minimum during training, can be calculated by performing projection to a given image representation base, |xyzo>, i.e., by performing a measurement on the state |Vfiltered>In the formula, VXCATk is not projected to a base because it is not in a quantum representation but in a classical voxel-type one; because it is transformed into quantum representation before the forward projection module 300, before said module it is still available in this form. Thus, it can be substituted in this form, while ^ / filtered originating from VXCATk (in the argument of the function f the matching VXCATand vfilteredare included) - because due to having been passed through the modules it is represented as a ket vector - can be substituted in the loss function specified for the classical case by the formula before the present quantum formula, based on the CNR formula by projecting it to a base. Optionally, we could also perform m measurements for each of the k images, in which case obviously k, m would stand under the summation symbol. Performing m measurements is advantageous because the quantum representation is stochastic, so performing the operations multiple times (this can be done because the quantum representation is fast) different results can be obtained. In such a way, variability (dispersion) is also obtained, which is also an advantage with respect to the classical approaches.However, in the case of full quantum representation, co-training the FP-NN and REC-NN quantum neural networks can be performed applying algorithmsdeveloped for the co-training of quantum neurons, e.g. the so-called quantum backpropagation algorithm (Beer, K., Bondarenko, D., Farrelly, T. et al.: Training deep quantum neural networks. A / afure Communications 11 , 808 (2020)) or other algorithms that make use of the specificities of quantum networks (e.g. quantum annealing, see for example Arnab Das and B. Chakrabarti: Quantum annealing and related optimization methods, 2005, Springer-Verlag Berlin / Heidelberg). In case of a hybrid quantum representation, the method of adjusting neuron weights is conventional, for example it is of backpropagation-gradient type.In stage 5 carried out optionally after stage 4, in-situ post-training is performed, wherein image synthesis (reconstruction) is made more accurate by the current measurement.In a preferred embodiment of the invention, therefore, after stage 4 the output is made more accurate by means of the current measurement. An important consideration, which leads to the advantage of optionally including stage 5, is that according to this embodiment, post-training is performed after specifying a given medical field in stage 4. Thereby it can be provided even more that the filtered 3D and 4D volumes maintain the accuracy according to a measure that is diagnostically relevant from the aspect of the given medical field. We are not aware of any previous application of in-situ post-training in such a framework.In stage 5, loss calculation is performed in the data representation space of the measurement, i.e., for example in the case of SPECT, projection images are compared, while in the case of CT and PET, respectively, sinograms (a sinogram is a type of representation in which projection data are arranged according to the angle of the projection; it is also meant to be included in the concept of “image”, or more generally, “visualising data package”) and LOR-s are compared (applying the reproduction-appropriateness loss function, for the implementation possibilities of this function see above).In Fig. 5, the post-training input pmeasuredmedical image 500 is the individual measurement, i.e., it originates from an actual investigation (examination). For that, a reconstruction is generated with the system according to the invention, with theinformation carried by the pmeasuredmedical image 500 being utilised for post-training prior to that.In the post-training process, the neural network REC-NN that implements the reconstruction module 200 reconstructs the post-training input pmeasuredmedical image 500 and generates the reconstructed vfilteredvolumetric image 530 as an output. For post-training, the reconstructed vfilteredvolumetric image is also back- projected applying the forward projection module 300 (FP-NN network), generating a post-training loss-check medical image 535 (Pfiltereddata), which is compared with the post-training input pmeasuredmedical image 500 (see the double arrow between them that in this example corresponds to the reproduction-appropriateness loss function). The reconstruction module 200 and the forward projection module 300 are provided with CT anatomy information 525 corresponding to the respective posttraining input pmeasuredmedical images 500 (this information is also to be measured or is to be available).In stage 5, it is preferable to apply a likelihood function for loss calculation, i.e. , it is preferable to determine the probability of pmeasuredoriginating from Pfiltered. it is this loss, i.e., the reproduction-appropriateness loss function that is to be maximised, or the negative logarithm thereof to be minimised (see for example the Poisson and SSIM-based approaches above), and we would like to modify the parameters of the reconstruction module 200 and the forward projection module 300 (REC-NN and FP-NN) accordingly, i.e., perform a post-training on those, typically only performing a few iteration cycles (preferably 1 -1000, particularly preferably 1 -300, even more preferably 1 -100, by way of example 2-8 or 2-4 iteration cycles), performing iteration also until improvement is detected in the values of the loss function. Due to the nature of the loss function, primarily such functions are established as loss functions of which the minimum is to be found (i.e., the losses are to be optimised). However, in an equivalent manner, based on the methodology of extremum search, a function to be maximised may also be chosen.Thus, in stage 5 an optimum is searched moving in a subspace, and accordingly, we expect that the extremum can be found earlier. When stage 5 is performed and iteration is applied, we would preferably also like to ensure that the results of stage 4 do not deteriorate (i.e., quantitative accuracy is provided). This can preferably bedone in two ways. Either by performing a few iterations according to stage 5, and then checking for any improvement, i.e. checking if we have left the vicinity of the optimum, or by still taking into consideration stage 4 as a constraint.According to the first option, for example a likelihood loss is calculated in stage 5 between the measured image and the generated one (i.e., Pfilteredmedical image, in general visualising data package), and the weights of the neural network are adjusted such that the negative log likelihood value applied as the loss is reduced. Thereafter, following one iteration or a few iterations, the (e.g. XCAT) training data (for which the location of reduction / accumulation is known) applied in stage 4 are passed through the forward projection and reconstruction modules trained with the post-training according to stage 5 (calculation of stage 4, i.e., the modules subjected to post-training are checked using stage 4 data) and it is checked whether the loss has been reduced also here. If it has not, the gradient is projected backward in the post-trained modules, training them according to stage 4, which is followed by performing stage 5 again. Therefore, in such a way stage 5, stage 4, and again stage 5 are performed alternately to search for an optimum. This process can be rendered more stable by applying numerical methods, for example the annealing (cooling) method that has already been mentioned.The other option is applying in stage 5, in a back-propagation algorithm, a constraint (boundary condition) that must be adhered to. This boundary condition is that stage 4 loss must not increase during post-training of the reconstruction and forward projection modules. That is, only steps along gradients with such directions are allowed in which also the other (stage 4) loss does not increase (training is performed in a way, and, thus, with this condition). For applying this constraint, for example the approach of neural conditional random field (NCRF) deep learning framework can be chosen (see Yi Li et al.: Cancer Metastasis Detection With Neural Conditional Random Field, arXiv:1806.07064v1 , 2018).For the case wherein neither of the options produces a result, such a criterion can also be specified (requiring certain parameters, e.g., the maximum number of calling stage 4 in the option one, or, if the boundary condition cannot be met through training in any direction in the option two) which, if the criterion is met, results in stopping the process according to both options, and in producing the output of thereconstruction module applying the system trained in stage 4 (that is, not modifying it in stage 5) in the same way as in the embodiments not including post-training (cf. the label “output” in Fig. 4). Besides that, in the present application the usual guidelines customarily applied in machine learning are applicable for the training (i.e., for when to stop it and declare it completed).For each of subsequent measurements (i.e. for scans related to new patients), pretraining is carried out for the same module state, always “resetting to zero”, i.e., removing the effects of previous measurements on them. After performing posttraining with post-training input pmeasuredmedical images 500, we return to a state of the reconstruction module 200 and the forward projection module 300 wherein stage 5 was started, i.e., for reconstructing the measurements of a subsequent patient, stage 5 post-training corresponding to the previous patient is not taken into consideration. Accordingly, it can be trusted that the training performed in stage 5 is based on appropriate grounds.Stage 5 is thus performed at the site of the examination where the trained system (apparatus) was installed for performing reconstruction. In stage 5 an actual measurement is applied, i.e., training is carried out for the actual measured image (this is online-type training), the system is trained with the image, and in the posttrained state of the system the output is generated with the reconstruction module.Thus, in addition to training, Fig. 5 also illustrates the application of the system, because following the completion of the post-training process involving a few iterations, the reconstructed filtered Vfilteredvolumetric image 530 will play the role of the output.In a quantum representation, in the case of quantum neural networks the pmeasuredimage is encoded in a quantum state |pmeasured> that constitutes the input of the REC-NN (quantum) operator (which is the quantum implementation of the reconstruction module 200). At the same time, in case of in-situ training, the actual measured image is applied for determining the loss, that is, the output quantum state |pfiitered> ofOperator FP-NN projected to a given image representation base is applied for comparison in a real space (and the obtained loss is applied for training the reconstruction module 200 and the forward projection module 300). Becausethe measured image is not in quantum representation, a measurement must be performed on the state |pfi|tered>iwhich produces an actual image (which images are different for different measurements), and a loss, for example L2, is calculated between the image measured with the apparatus and the image obtained from the quantum measurement. It is this loss that is to be minimised (or more generally, to find its extremum, it can also be maximalized, if it is so constructed), i.e. , we adjust the parameters of the quantum neural networks. After training, the state |Vfiltered> is generated again applying the post-trained reconstruction module 200, and the required reconstructed image (Vq-out) can be obtained by projecting this state to the image space:This formula includes the projection of the state |Vfiltered> to a base, that is, the result is obtained such that the reconstruction module 200 (subjected to pre-training) generates the state | vfiltered>, which is then projected to the base.It is however also possible to transcode the measured image into qubits, and to assess the similarity of the images encoded in a quantum image representation, for example by performing a quantum measurement on both (see Fei Yan et al., "Assessing the similarity of quantum images based on probability measurements," 2012 IEEE Congress on Evolutionary Computation, Brisbane, QLD, 2012, pp. 1 -6.).In the case of quantum neural networks, the inputs designated by CT (CT anatomy information 125, 425 and 525) are also represented in qubits, and appear as secondary, supplementary inputs of the quantum operators REC-NN and FP-NN.Compared to the classical representation, the implementation described above preferably provides increased speed, because the quantum implementation in inherently faster. This precisely one of the motivations for developing quantum computers. Although in-situ post-training is highly preferable, it is often not possible to wait for the results for a prolonged time (for example to wait for stage 5 to produce the output). In a quantum representation, this is much more feasible.It is noted that a possible approach of implementing the five-stage system described above - i.e., an embodiment of the invention - is for example the so-called “data-driven quantum circuit learning” (DDQCL; Marcello Benedetti et al.: A generative modelling approach for benchmarking and training shallow quantum circuits, Nature, npj Quantum Information (2019) 5:45; see hybrid quantum representation above). In this approach, a complex, generic network is constituted by generic gates producing phase shift of qubits representing the image and by gates bringing the qubits to entangled state. The individual phase shift values of the quantum gates are tuneable parameters that constitute the weights of a neural network; by training the neural network we search for an appropriate setting of the weights. The output of the classical neural network trained applying the hybrid quantum representation is therefore not an output image, but the weights of these will determine the operation of the series of gates implementing quantum image processing. The image reconstruction operation is in turn carried out by the quantum computer.To sum up the above it is noted that the two neural networks (REC-NN and FP-NN) trained in stages 2-5 (thus, preferably also applying the optional stage 5), the training process, and the data prepared for them in stage 1 collectively define an extremely preferred embodiment (with an appropriately prepared training database: stage 1 , and considering actual in-situ individual data: stage 5).In Fig. 6A, the embodiment of Figs. 1 -5 is summarised in a schematic flow diagram. Thanks to its schematic nature, Fig. 6A depicts a more schematic embodiment than the one of Figs. 1 -5, for example stage 1 does not appear at all, and of course many specifics of stages 2-5 are also not included (however, the optional stage 5 is included, see also below).In Fig. 6A, arrows starting from the blocks corresponding to the various trainings (cf. the notations of these with Figs. 2-5) indicate the respective module trained applying the given training. To sum up, the first pre-training corresponding to operational step S220 is applied for training the reconstruction module 200, while in operational step S330 the forward projection module 300 is trained. Operational steps S440, S450, S540 and S550 are included in stages 4 and 5, i.e., in the medical field-specific training step and the post-training step; these steps are applied for training both the reconstruction module 200 and the forward projection module 300.In the embodiment illustrated in Figs. 7-11 , in contrast to the modality-preserving embodiment illustrated in Figs. 1 -5, such a system is applied which is adapted for a modality-switching reconstruction of a medical visualising data package (e.g., a medical image; for example, we start from a SPECT or PET image, from which we generate a synthetic CT image that is of course not equivalent to a real CT image but is excellently suitable for certain purposes, such as for absorption correction or scattering correction; such a modality switching is illustrated in Figs. 7-11 ), the reconstruction module is configured as a modality-switching reconstruction module adapted for generating a modality-switching volumetric visualising data package by processing a medical visualising data package. The term of “the system is adapted for a modality-switching reconstruction of a medical visualising data package” also implies that the forward projection module applied in the training method performs modality-switching in the reverse direction with respect of the reconstruction module. The volumetric visualising data package processed by the forward projection module will be subjected to modality switching in an opposite (inverse) way relative to the medical visualising data package processed by the reconstruction module. The descriptions (e.g. terminology) related to the two - i.e., modality-preserving and modality-switching - embodiments, also due to their analogy, are applicable to both, unless considerations arising from their differences prevent this.Thus, the invention in an embodiment provides a novel solution for implementing modality switching, thereby opening up a possibility for gathering data corresponding to modalities not available on the level of the scan (e.g. gathering information related to CT that can only be generated applying costly apparatus; see also below in relation to this). There are known techniques for implementing modality switching (see e.g. T. Huynh et al., "Estimating CT Image From MRI Data Using Structured Random Forest and Auto-Context Model," in IEEE Transactions on Medical Imaging, vol. 35, no. 1 , pp. 174-183, Jan. 2016), however, the relevant embodiment of the invention provides a solution based on different grounds.The basic conception of the embodiment of Figs. 7-11 is identical to that of the embodiment of Figs. 1 -5; training of the system is performed following analogous principles, and the modules with correspondent functionality contribute to thecommon foundations of the two embodiments, at the same time - in accordance with the databases used for their training - the embodiment of Figs. 1 -5 preserves modality during reconstruction, while the embodiment of Figs. 7-11 switches the modality then. However, this affects only the contents of the database used for training (it will behave corresponding to the contents of the images - in general: visualising data packages - it was trained with) i.e. it does not affect the basic principles applied in the invention.Similarly to Fig. 1 , Fig. 7 illustrates the preparation of the training database, i.e., the generation of training data. In the embodiment illustrated in Fig. 7, transmodal training data subsets are applied, that is, such training images and CT data on the basis of which latter the reconstructions corresponding to the specific training images can be constructed. In general terms, a respective modality-switching medical visualising data package (in the present example, see a medical image 625 as a modality-switching medical visualising data package that in the present embodiment has CT modality corresponding to the base SPECT or PET modality) corresponds to each given medical visualising data package of the base set (in the present example, medical images of a base set 610). In this way, each reconstructed volumetric image of a second auxiliary pre-training set 620 will correspond to a respective medical image of the base set 610, and the training database is constructed for pre-training applying these matching data.In accordance with Fig. 7, a splitter module 605 is applied for generating measurements with shorter measurement time / lower dose from PHmedical images of the base set 610, generating j measurements from each image (these will be PLj medical images of the first auxiliary pre-training set 615). Thereafter, the reconstruction module 600 (labelled “CT rec” in the figure) is applied for generating from the medical images 625 the volumetric images of the second auxiliary pretraining set 620. Thus, the latter will be the reconstruction corresponding to the PHSPECT recording, such that, as it is illustrated in Fig. 7, a direct reconstruction of the PHSPECT recording is not done.If, therefore, stage 1 is preferably applied in modality-switching embodiments, then, in such an embodiment (as illustrated by the embodiment shown in Fig. 7) preferably- in the first auxiliary pre-training set of the medical visualising data packages, pretraining auxiliary medical visualising data package groups comprising medical visualising data packages with different noise levels (noise) each generated from a base set of medical visualising data packages with a splitter module are applied, and, in the second pre-training step, as visualisation information corresponding to respective given medical visualising data packages of the first auxiliary pre-training set of the medical visualising data packages, such medical visualising data packages of the base set of the medical visualising data packages are applied that correspond to each of the medical visualising data packages of the first auxiliary pre-training set of the medical visualising data packages (see the upper row in Fig. 7 indicating that in this embodiment the base set 610 is applied, from which the first auxiliary pre-training set 615 is generated), and- the second auxiliary pre-training set of the volumetric visualising data packages corresponding to the first auxiliary pre-training set of the medical visualising data packages is generated by processing the modality-switching medical visualising data packages corresponding to each of the medical visualising data packages of the base set of the medical visualising data packages by means of a different- from-machine learning-based reconstruction module (see the bottom row in Fig. 7 indicating that in the present embodiment for example the medical images 625 are applied as modality-switching medical visualising data packages, from which a second auxiliary pre-training set 620 is generated).In the present embodiment, therefore, the volumetric visualising data packages of the second auxiliary pre-training set 620 are generated, with different-from-machine learning-based module, not on the basis of the base set (but still for the visualising data packages thereof, because they will be mutually corresponding) - unlike in the case illustrated in Fig. 1 - but based on the modality-switching medical visualising data packages (that is, visualising data packages having a different modality with respect to the medical visualising data packages of the base set, i.e. , having the modality to which we are switching). Thus, in this embodiment the training data are prepared in this way, i.e., these modality-switching medical visualising data packages must also be available.In this case, by referring to the second auxiliary pre-training set of the volumetric visualising data packages corresponding to the first auxiliary pre-training set of medical visualising data packages it is therefore meant that the volumetric visualising data packages of the second auxiliary pre-training set 620 correspond to medical visualising data packages of the first auxiliary pre-training set 615 generated from the base set 610, because said volumetric visualising data packages have been generated through reconstruction from medical images 625 (generally: modality-switching medical visualising data packages) each corresponding to respective elements of the base set 610. It can be understood that, applying the splitter index j, in particular, visualising data packages originating from medical visualising data packages of the same base set 610 correspond to each other (i.e., there is a one-to-one mapping / correspondence defined between the elements of the base set 610 and the medical images 625). These mutual correspondences are therefore determined also in this case.In this embodiment, furthermore, preferably (this option is not shown in Fig. 7 for the case of switching from SPECT or PET to CT modality, because in CT reconstruction the use of such parameters is not typical, but in this case and also in other modalityswitching cases it is possible to apply such, in this case the parameters would be used in the splitter module which is here - a splitter module 600 in Fig. 7; applying the present embodiment, in Fig. 8 a comparison according to a parameter p would be incorporated in the loss, like in Fig. 2)- for each of the medical visualising data packages of the base set of the medical visualising data packages, a plurality of volumetric visualising data packages of a second auxiliary pre-training set are generated by processing, by means of a different-from -machine learning-based reconstruction module, the modalityswitching medical visualising data packages with a plurality of parameter values different from each other of reconstruction parameters corresponding to the different-from -machine learning-based reconstruction module, and- the first pre-training loss function is composed based on a plurality of volumetric visualising data packages of the second auxiliary pre-training set generated by processing with a plurality of parameter values different from each other of the reconstruction parameter and on first pre-training volumetric visualising datapackages generated by processing the medical visualising data packages of the first auxiliary pre-training set by means of the reconstruction module.In Fig. 8, pre-training of the reconstruction module 700 is performed analogously to Fig. 2 (like in Fig. 2, the module is also designated REC-NN, indicating that it is preferably implemented by means of a neural network). In this embodiment, the reconstruction module 700 generates from the PLj medical images (for example, SPECT projection images) of the first auxiliary pre-training set 615 vCTfilteredj volumetric images 710 (i.e., pseudo-CT images). In the case of supervised training illustrated in Fig. 8 (an operational step S720 corresponding to the pre-training of the reconstruction module 700 starts from a double arrow), the loss function measures similarity to the original CT image (which forms a part of the training database, see Fig. 7), and the parameters of the reconstruction module 700 (RECNN) are set by minimising this loss.In relation to Fig. 8 it is important to note that it is analogous to Fig. 2, with the difference that, because the training database contains modality-switching pairs (and modality-preserving pairs in case of Fig. 2), during pre-training the reconstruction module 700 is trained for modality-switching reconstruction (and is also trained further in such a way during medical field-specific training, see Fig. 10). Of course, the forward projection module 800 (designated in the figure as FP-NN according to the corresponding neural network) is also pre-trained in line with these principles, as illustrated in Fig. 9.The operator FP-NN (that is, the forward projection neural network) implementing the forward projection module 800 is also a transmodal operator, i.e., it gives an estimate made from the VCTvolumetric image to the corresponding SPECT image. In Fig. 9, analogously to Fig. 3, the input of the reconstruction module 700 is constituted by the PLj medical images of the first auxiliary pre-training set 615, from which it generates respective filtered vCT filteredvolumetric images 810. From them, the forward projection module 800 generates the (second pre-training) pfilteredj medical images 820, each of which it compares with a respective corresponding PHmedical image of the base set 610 during training (operational step S830 corresponds to this). This is a supervised training, where the loss function means a comparison of the Pfilteredj medical images 820 and the PHmedical images of thebase set 610 (e.g. utilising L1 or L2 distance, or some kind of reproductionappropriateness loss applied also in stage 5, for example likelihood loss, for comparing projection images the latter can be a good choice; the PHsame medical image of the base set 610 corresponds to more than one Pfilteredj).Fig. 10 shows stage 4 in the modality-switching embodiment that is also analogous to the modality-preserving embodiment illustrated in Fig. 4. Fig. 10 thus illustrates the training of the medical-specific training - which can also be called post-training - of the reconstruction module 700 designated as REC-NN and the forward projection module 800 designated as FP-NN.Accordingly, in stage 4, starting from volumetric images of a medical field-specific training set 910, medical field-specific training medical images 920 forward- projected by the forward projection module 800 are generated, which are returned to the reconstruction module 700 that generates reconstructed medical field-specific training volumetric images 930 based on them (later on, in the trained state of the system these reconstructed volumetric images can function as the output, if the optional stage 5 is not performed). For training, matching volumetric images of the medical field-specific training set 910 and volumetric images 930 are compared (see the considerations set forth above in relation to Fig. 4, and also see below), and based on the result of the comparison, that is, based on the loss calculated from it, the reconstruction module 700 is trained in operational step S940 and the forward projection module 800 is trained in operational step S950.To sum up, therefore, a volumetric image obtained going around the system is compared with an original one in line with appropriate principles (either the principles described in relation to Fig. 4 or according to those described below).As usual, the training process is performed in multiple cycles applying the volumetric images of the medical field-specific training set 910, in the course of which the machine learning-based reconstruction module 700 and forward projection module 800 are trained after each cycle, i.e., these modules learn continuously and get better and better due to machine learning. Training essentially involves the optimalisation of the machine learning modules (such that they generate the best possible results), typically by minimising the loss function.In the case of transmodality, i.e., modality switching, therefore, there are no significant differences with respect to the modality-preserving embodiment.To the above, let us consider an example in which the medical field is related to cardiac SPECT scans. Then, the two SPECT - modality-switching - reconstructions are the pseudo-CT image (volumetric image 930, VCTin Fig. 10) and the image of a mathematical phantom (a vXCAT’CTvolumetric image of the medical field-specific training set 910). In these, for example the amount of reduction in the cardiac wall of the left ventricle can be compared, or another evaluation approach can be applied. According to the modality-switching in Fig. 7-11 , the vXCAT-CTvolumetric images of the medical field-specific training set 910 have CT modality (hence the designation in the index).In this case, the object of further training is that there should be no significant difference between the modality-switching reconstructions of the two SPECTs.For comparing the volumetric visualising data packages (for example volumetric images) to be compared, a deviation evaluator module for visualising data packages can be utilised (e.g. a cardiology or clinical (deviation) evaluator - in certain medical fields - or, somewhat more generally, a medical image evaluator). These can be for example implemented as the Emory Cardiac toolbox described in an article by Prieto Canalejo, M.A. et al: Synthetic Attenuation Correction Maps for SPECT Imaging Using Deep Learning: A Study on Myocardial Perfusion Imaging, Diagnostics 2023, 13, 2214., a dedicated approach made by the Cedars company (i.e. dedicated software applications, see Guido Germano et al.: Quantitation in gated perfusion SPECT imaging: The Cedars-Sinai approach, Journal of Nuclear Cardiology, Volume 14, Issue 4, 2007, Pages 433-454), a 4DM framework similar to the Cedars approach (see Ficaro, E.P., Lee, B.C., Kritzman, J.N. et al., Corridor4DM: The Michigan method for quantitative nuclear cardiology. J Nucl Cardiol 14, 455-465 (2007)), or an equivalent cardiological evaluator.In the approaches applied in the articles referenced above, the images input to the deviation evaluator module are measured against a “normal” database (which comprises healthy data), that is, in case the matching VXCATand Vfilteredvolumetric images when fed to the difference evaluator module, it will measure them againstthis (the available - external - evaluators have not been developed for neural network training). Thereafter, we evaluate the “distance” of these two matching results (deviation data), for example applying L1 or L2 metrics, thus obtaining the corresponding loss contribution, and then, by running the stage 4 for the set of VXCATvolumetric images the complete loss value is obtained (i.e. , the contribution of the model reproduction loss function corresponding to stage 4).In this embodiment, therefore, the first deviation data and second deviation data expediently applied for determining the model reproduction loss function are determined applying a deviation evaluator module (in accordance with the previous paragraph, the model reproduction loss function is determined applying their deviation, for example in the case of data represented in voxels, by adding up per- voxel deviation in some way).In an example implemented according to the article referred to above describing the Emory Cardiac toolbox, the deviation evaluator module performs the following: it generates a score for the reduction in the left ventricle (i.e. what extent of reduction it is classified), and also generates a correction map that indicates the extent of the cardiac wall injury and the extent of the ischaemic cardiac disease (this will be compared for the VXCATand Vfilteredvolumetric images). In another similar example, e.g. in case of cardiology the deviation evaluator module may for example expect a so-called “stress” image and a “rest” image and may compare the two, as cardiac stress testing (stress / rest) is an established clinical protocol. In the invention this approach can be made use of by inputting the two images, that is the VXCATand Vfiitered volumetric images instead the “stress” and “rest” images. The loss can be derived from the output data indicative of cardiac health also in this case.Therefore, it seems to be expedient to apply the deviation evaluator module for loss determination especially in the modality-switching embodiments, but it can also be applied in the modality-preserving embodiments (i.e. in the embodiment illustrated in Figs. 1 -5), and the loss calculation method described above for stage 4 in relation to the modality-preserving embodiments can also be applied in the modalityswitching embodiments (in both cases, volumetric feature is compared to other volumetric feature).Thus, the application of the deviation evaluator module is not necessary, it can be a preferable solution basically in case the external evaluation software applied therein is fast enough for training (during training, the evaluator module must be called very frequently, so if it does not have the required speed, training may be impossible to perform). Therefore, if the required speed cannot be provided, it could be expedient to apply the approach described in relation to Fig. 4 (with respect to the modalitypreserving embodiments) also in the present modality-switching stage 4.Like Fig. 5, Fig. 11 also illustrates the in-situ training process (stage 5), where the networks are adapted to specific SPECT measurements to ensure the consistency of the measurements. Fig. 11 is also analogous to Fig. 5, according to which in a modality-switching case the optional stage 5 is carried out as follows. We start from the post-training input medical image 1000 that has been recorded in-situ. This is fed to the reconstruction module 700 (that has been put through the previous stages, including stage 4), which generates a post-training volumetric image 1030 that can also function as an output after the completion of stage 5 training. This is passed on to the forward projection module 800 that generates a post-training loss-check medical image 1035 based on it. The post-training loss-check medical image 1035 is compared with the post-training input medical image 1000, and in operational steps S1040 and S1050, respectively, in-situ training of the reconstruction module 700, and the forward projection module 800 is performed based on this. Otherwise, the stage 5 training process is performed analogously to what has been described above in relation to Fig. 5.In the case of the illustrated modality switching, at the end of the process such a pseudo-CT image is generated from the exemplary SPECT measurement (of course other types of modality switching can also be applied) that is optimal for the given medical field (for the given medical application) - e.g. for supporting SPECT reconstruction (e.g. absorption correction or scattering correction) - and is consistent with the actual measurement. The pseudo-CT image thereby obtained is therefore not necessarily suitable for a physician to establish a diagnosis by analysing anatomical conditions.This is because SPECT measures different features than CT, and at a different resolution. Because of that, there will be such features (e.g. calcified coronaries ofthe heart) that cannot possibly be included in a CT image generated from a SPECT scan through modality switching.Furthermore, a SPECT scan and a corresponding CT will always scan the patient at different positions and with different parameters. With SPECT, the measurement may take for example 20 minutes, and the breathing movements are integrated. In contrast to that, chest CT scans are produced over a short time interval (for example about 10 seconds), with the patient typically holding his / her breath. As a result of this, for example a lung tumour is at a different position in the CT than in the SPECT. This causes significant differences when the two are compared. Such differences are eliminated by means of the CT image obtained through modality switching (as it has no motion blurring error). At the same time, in accordance with the information content of the starting modality, it can contain different information from the start than an original CT scan. It is however not one of our objectives to remedy or avoid this as the modality switching is not used for that.In principle, the modality-switching embodiment can be applied for performing any type of modality-switching reconstruction, but basically it is expedient to perform the followings: MRI^CT; SPECT^CT; SPECT^MRI; PET^CT; PET->MRI.Any other modality switch is less preferable, i.e. expediently we do not start from the utilisation of a costly apparatus that causes radiation exposure just in order to generate such a reconstruction from it through modality switching that corresponds to the modality of a more favourably applicable device. Preferably, the application of MRI and CT apparatus as costly components and, in the case of CT it is also a device with significant radiation exposure, can be replaced by modality-switching reconstruction. However, switching from MRI to CT modality is also preferable because the application of MRI is more expedient, and a CT reconstruction may also be obtained as a result in this way. Therefore, following the ALARA (as low as reasonable) principle, if sufficient information can be obtained already from a SPECT (i.e., without CT) for absorption / scattering correction and for simple anatomical localisation, then it is expedient to avoid additional radiation exposure of the patient.Thus, in certain cases, for example, a CT apparatus may not be necessary (its investment costs can be saved) and without applying additional radiation exposure, a CT result that allows for performing localisation and activity distribution evaluation can still be obtained. Preferably, motion blur artefacts can also be reduced, because e.g. we start from SPECT, i.e. better results can also be obtained in 4D. In light of the above, it is preferable to apply such modality switching wherein the end modality is CT or MRI irrespective of the starting modality that can be any of the applicable modalities.In the following, further considerations and interpretations related to a topic in connection with the invention are provided.The so-called “transformers” are neural networks originally devised for text processing (Neural Language Processing, NLP) in mind. For example, if we would like to translate a sentence for Hungarian into English, it is obviously not sufficient to replace individual words with their respective English translations. Moreover, in many cases it is also expedient to consider the broader context in order to find the right expression or its proper usage. The transformers are such neural network solutions that are also able to the learn the proper relationships between words.This idea has been introduced also into image processing, because, similarly to texts, images are also “not noise”, i.e., they are not sets of unrelated values. If an image is decomposed into patches (pieces), they will behave similarly to words. In image processing, such networks are referred to as Visual Transformers (A. Dosovitskiy et al.: An image is worth 16x16 words: transformers for image recognition at scale, arXiv:2010.11929v2, 2021 ; see also e.g. F. Shamshad et al.: Transformers in Medical Imaging: A Survey, arXiv:2201 ,09873v1 , 2022).In relation to the invention, the following can be set forth: we have an input measured image (e.g. sinogram), and we know what the output (reconstructed image) should be. The transformer learns a matrix W. In case the task is image reconstruction, then the matrix W, which is also called Self-Attention, will essentially be the system matrix encoding the mapping, i.e., encoding the physics, which specifies the relationship between the input and the output. The “words” are of course not necessarily the measured pixels, it may also be that the measured sinograms are assigned to thelocal pixels of the reconstructed image by decomposing them into sine waves (in case of quantum neural networks, this probability space will be encoded in a few qubits).Visual transformers have the same difficulty as end-to-end with other approaches based on neural networks (“unrolled” type approaches typically do not implement an end-to-end reconstruction): in order that they can generate sufficiently accurate reconstructed images, they must typically be trained on millions of images. That is, the internal data representation of the Visual Transformer typically encompasses a very large latent space (this is referred to as “diversity”).Stages 4 and 5 of the invention are intended precisely for providing that the SelfAttention matrix of such a pre-trained Visual Transformer, i.e. , the system matrix, is focused on a particular application (medical field), and that it generates an accurate output for the given measurement. Thus, stages 4 and 5 can be considered as a special solution for making a Visual Transformer neural network more accurate. This is necessary because images are expensive in medical imaging (cf.: in the case of self-learning cars thirty images are generated every second; in a given hospital, in a given medical field, a day or even a month may be needed to produce as many images). Therefore, the invention may also be considered as a training method for eliminating the problem that the number of available images is not sufficient for ensuring that the output of the neural network is accurate.In the following, the invention is differentiated from certain prior art approaches.In paragraph

[0191] of US 2022 / 0101526 A1 mentioned in the introduction it is set forth that certain parameters of the imaging - e.g. the current and voltage of the X- ray tube - that fundamentally determine the contents of the generated image can be treated as parameters of the image processing neural network. Conversely, according to the invention, preferably in stage 1 the database of VHPreconstructed images is generated applying the reconstruction parameter set p, which is applied in stage 2 for training the neural network (that typically has the characteristics of an autoencoder), which is preferably utilised in that stage which typically does not include the reconstruction parameter p as an input parameter, but is applied in sucha manner for specifying a range, i.e. for preparing the training database such that the reconstruction module is trained to be applicable more broadly.The above-described generic concept based on fixing the parameters is included in US 2018 / 0360313 A1 , wherein in paragraphs

[0034] -

[0035] it is specified that deep learning is performed for the specific application, and that the training data vary depending on the specific application. This document also indicates that fixing of the parameters, i.e., performing training from the outset for a specific application is the generally applied approach according to the prior art.In US 11 ,244,480 B2 a reconstruction approach based on deep neural networks is disclosed, wherein a mathematical phantom (e.g. XCAT) and projection means accurately modelling the imaging (e.g. the Monte-Carlo based GATE [Geant4 Application for Emission Tomography]) are applied for generating the training database. By generating the database (with this), the neural network performing the reconstruction can be trained.The present invention significantly differs from this, because in the invention the system is trained on real measured images and their reconstructions, that is, incorporates the unique distortions of the imaging system. In addition to that, in our case, in stage 4, through the introduction of a medical field-specific loss, the artificial investigation area model (e.g. mathematical phantom) is applied for making more accurate (i.e. train further) the already pre-trained forward and backward projection neural networks preferably applied for implementing the reconstruction and forward projection modules. This step, i.e. the application of stage 4 is intended for enabling the system to find the solution that is optimal, most favourable for the given medical field. In other words, the latent space containing the reconstruction structures evolves during the training in such a way that a reconstruction parameter setting that is optimal for the given medical field is also established, that is, our objective with further training is to achieve that the neural networks have such reconstruction parameters that the reconstruction is optimal for the given medical field.In US 2022 / 0327665 A1 , a neural network is utilised for improving the imaging performance of an imaging system; for example, in the case of a TOF-PET, starting from a system with a temporal resolution of 800 ps a virtual imaging systemequalling a system with a temporal resolution of 250 ps is created. Image enhancement is implemented by modelling the imaging systems having different imaging properties, and training the neural network on the image pairs generated in such a way (that is, a modelled training database is applied). This way such a deep network is trained that can be expected to be able to improve the imaging parameters, such as the resolution, of the real imaging system.In contrast to that, the invention is based on a pre-trained system that is made more accurate utilising a medical field-specific loss. That is, according to the invention we do not aim at improving specific imaging properties of the camera but would rather like to generate as accurate a reconstruction as possible while taking into consideration the specificities of the given medical field, utilising a given image having a given information content, i.e., the invention is fundamentally different from the concept of US 2022 / 0327665 A1 .According to US 2020 / 0027530 A1 , patient data are simulated on the basis of the data of real patients, that is, models are generated for training the artificial intelligence algorithms, and the algorithms are trained directly using the models.In contrast to that, the invention is based on a real database where the “ground truth” is not known (i.e. we have only an estimate of the activity distribution, see e.g. its calculation in stage 1 ), and the artificial (e.g. mathematical) models (e.g. anthropomorphic phantoms) are only applied for ensuring that the pre-trained system can select a reconstruction setting that is optimal for a given medical field. That is, this concept is fundamentally different from the known approach. Based on the known approaches, training performed exclusively utilising model data has not proven to be successful. However, having overcome the professional prejudice based on this experience, the invention still applies such model data in stage 4 under appropriate circumstances, i.e. in a way that the disadvantages of these can be suppressed while exploiting their advantages.Compared to the articles describing approaches to reconstruction listed in the introduction (A. Mehranian et al.: Model-Based Deep Learning PET Image Reconstruction Using Forward-Backward Splitting Expectation-Maximization, IEEE Transactions on Radiation and Plasma Medical Sciences, Vol. 5, Issue: 1 , pp.54-64, January 2021 ; K. Gong et al.: Iterative PET Image Reconstruction Using Convolutional Neural Network Representation, IEEE Trans Med Imaging, 38(3): 675-685, Mar 2019; J. Adler, 0. Oktem: Learned Primer-dual Reconstruction, arXiv:1707.06474v3), an important difference in the invention is that the systems described in the articles are intended for reconstruction, and inside this - typically for reducing the amount of required training data - backward projection can be applied. In contrast to that, in the invention separate reconstruction and forward projection modules (operators) are applied that are trained in stages 2 and 3. In stages 2 and 3, training is performed (according to Figs. 2 and 3) by first performing the training (i.e. , pre-training) of the reconstruction module 200, and then the already trained reconstruction module 200 (that has undergone pre-training) takes part in training the forward projection module 300.Training the reconstruction module and the forward projection module in such two separate stages does not appear in any of the prior documents, also in accordance with the fact that there the aim is to tune these modules “to be complete” for reconstruction, while in the case of the invention it is an express aim to train separate reconstruction and forward projection modules, because in stages 4 and 5 these modules will undergo further training separately.Closely related to that, in the prior documents the express aim of “tuning to be complete” is to narrow down the parameter set of the reconstruction operator by the time training is completed (because therein no further stages are applied following the disclosed training), while in the case of the invention our express intention is to “leave open” the parameters at the end of stage 3, i.e., we would like to obtain modules having been trained in this way by the end of stage 3 such that they can be effectively subjected to stages 4 and 5. In other words, our aim is to provide that the latent space is sufficiently broad at the end of stage 3 so that it can be narrowed in stages 4 and 5 in a preferable manner (for this narrowing it would be disadvantageous if the latent space had been previously narrowed down too much).Using alternative terminology, this can also be formulated as the approaches according to the prior documents apply more concentrated training, while the invention typically utilises a broader database compared to them. Also, as a result of this, after stages 2 and 3 the reconstruction according to the invention will typicallynot be at the level achieved by the prior approaches, i.e. , the reconstruction provided by the invention will be typically worse compared to them; however - and this is important to emphasize - after stages 4 and 5 it will be potentially much better than the reconstruction provided by the prior art approaches.The training sequence applied in the invention - according to which the separate reconstruction module and forward projection module are pre-trained in stages 2 and 3, which is followed in stage 4 by further training them with the help of artificial models, and, optionally, in stage 5 with the help of in-situ recorded measured data- is preferably able to achieve very high effectiveness. It is emphasised that, in the training sequence applied according to the invention, training with artificial models- for example phantoms - is applied for medical field-specific further training, i.e. fine-tuning, and not for performing training by using them exclusively, i.e. not for a priori training with these.Stage 5 receives an already trained system that is trained further applying a loss that is derived from a single image. In stage 5, therefore, the system that has been trained in stages 2-4 is trained further, i.e. previous training is already encoded in the parameters of the network. These parameters are fine-tuned in stage 5.When a new measurement image is received (from a new patient examination), we proceed again with the modules adjusted in stage 4 (i.e. we reset the parameters of the neural network), that is, the system does not undergo long-term training with the images of stage 5, in other words, information gathered from the previous patient will not affect the current one. This is necessary because otherwise the algorithm would be changing constantly, and there would be no possibility of control should it lead to a wrong result.Although the above-mentioned articles (Jianan Cui et al.: Population and individual information guided PET image denoising using deep neural network, Conference Paper, May 2019 and Jianan Cui et al.: PET image denoising using unsupervised deep learning, European Journal of Nuclear Medicine and Molecular Imaging, 2019 Dec; 46(13): 2780-2789) do include combining training with individual information, however in a completely different framework and in a different sequence than it is optionally included in the invention. In the latter document a system trained this wayis validated with the help of a so-called BrainWeb phantom, i.e. , post-verification is performed on the trained system with that to check to what degree it returns the results of the artificial (generated) BrainWeb phantom that are naturally known in advance.In contrast to that, in the invention further training using in-situ measurements in the optional stage 5 can be inserted after medical field-specific further training, and, particularly preferably, for further training a system comprising a reconstruction module and a forward projection module. A big advantage of this is that a - sometimes uncontrolled - in-situ further training is applied not in a generically trained system, but to control a system applying pre-trained reconstruction and forward projection modules and has been subjected to medical field-specific further training. Thanks to that, by applying stage 5, diagnostic goodness of the approach according to the invention can also be ensured effectively (cf. these considerations and the related points of the introduction).Besides without applying stage 5 the system does not get the control provided by measurements (that is, applying of stage 5 provides a “control” with respect to the system trained in a medical field-specific manner), it is also functional in such a case; having completed stage 4 the reconstruction module provides output as if stage 5 had also been performed on the system.Including stage 5 could be expedient for handling of patient-specific aspects (e.g. body weight, specificities of certain organs, the presence of a pacemaker or catheter). The stages of the invention can be taken to “mimic” the physician in that the parameters that are to be viewed / measured by the physician are tested (narrowing down to these in a medical field-specific manner) by also taking into consideration the measurement data of a large number of previous patients (this can be matched with the memory of the physician). Besides that, by optionally including stage 5, specificities of the given patient are preferably also taken into consideration.The following additional remarks are made to the two articles mentioned hereabove, particularly to the first one. These disclose a type of post-training. According to the approach described in the articles, two neural networks are trained. One of them,the network f1 is adapted for synthesising a PET image having high signal-to-noise ratio based on co-registered MRI images. This network is trained based on a coregistered PET+MRI database. The other network f2 is adapted for performing further training of the network synthesising PET images from MRI ones based on the parameters of the previously trained network, on the currently measured noisy PET image, and on the measured MRI image. That is, in some way the freshly measured PET image addresses, parametrises the MRI PET generative neural network.Conversely, in stage 5 preferably applied in the invention, the generative network is not parametrised by another modality, but instead a filtered projection image obtained through the chain REC-NN->FP-NN is compared to the measured one. Therefore, the REC-NN and FP-NN networks are trained further such that they operate as consistently as possible from the aspect of the given measurement: i.e., that one of them generates - to a good approximation - the measured projection image, which also results in automatically producing the reconstructed Vfilteredvolumetric image as an output. Hence, the invention applies post-training in a completely different way than what is described in the article, post-training being only an optional component of the invention, with the final aim of further fine-tuning the reconstruction in a medical field-specific manner, and not de-noising the image like in the article.In summary, in certain embodiments of the solution according to the invention:- it comprises multiple neural networks operating in a coordinated manner, i.e. such neural networks are applied therein;- these are trained using a special, multiple-stage training process;- the conventional neural networks can be preferably replaced by quantum neurons storing the images in quantum states (full quantum representation); while hybrid quantum representations are also conceivable (see the discussion above on quantum gates parametrised with classical neural networks), being established by combining the classical and full quantum representations (their implementation does not necessarily require quantum computers), that is, such a preferable hybrid solution is also feasible (in relation to quantum neural networks and their potential applications see Jeswal, S.K., Chakraverty, S.: Recent Developments andApplications in Quantum Neural Network: A Review. Arch Computat Methods Eng 26, 793-807 (2019)).The (image) reconstruction apparatus and the method performed with it preferably has the following advantageous features:- it makes use of the common, population-level information contained in similar examinations,- it automatically sets the required reconstruction parameters in a medical fieldspecific manner and in a way that is specific for the given individual measurement, and thus it maximises the diagnostic value of the quantitative examination,- it thereby preferably allows that the measurement time and the dose the patient is exposed to from X-ray or injected isotope can be significantly reduced (this can be achieved by including the optional stage 1 ), and- it preferably makes use of the individual information contained in the individual measurement (see the optional stage 5).In an embodiment, the invention can be considered as a medical field-optimised (medical application-optimised, i.e., trained for that) artificial intelligence- and preferably quantum neural network-based quantitative image reconstruction method (more generally, a reconstruction system, the reconstruction with which can be considered a reconstruction method), and a related training method.The invention is, of course, not limited to the preferred embodiments described in detail above, but further variants, modifications and developments are possible within the scope of protection determined by the claims.

Claims

CLAIMS1 . A training method for training a system adapted for reconstructing a medical visualising data package, wherein the system comprises a reconstruction module (200, 700) based on machine learning, adapted for generating a volumetric visualising data package by processing a medical visualising data package, and such a system is applied in course of the training method which comprises a forward projection module (300, 800) based on machine learning, adapted for generating a medical visualising data package by processing a volumetric visualising data package, and the training method comprising the following steps of:- generating, in a first pre-training step, a first pre-training volumetric visualising data package (210, 710) for each of those by processing a first auxiliary pre-training set (115, 615) of medical visualising data packages by means of the reconstruction module (200, 700), and performing a first pre-training (S220, S720) on the reconstruction module (200, 700) with a first pre-training loss function determined based on a second auxiliary pretraining set (120, 620) of volumetric visualising data packages corresponding to the first auxiliary pre-training set (115, 615) of the medical visualising data packages and on the first pre-training volumetric visualising data packages (210, 710),- generating, in a second pre-training step after the first pre-training step, a second pre-training volumetric visualising data package (310, 810) for each of those by processing the first auxiliary pre-training set (115, 615) of the medical visualising data packages by means of the reconstruction module (200, 700), generating a second pre-training medical visualising data package (320, 820) for each of those by processing each of the second pre-training volumetric visualising data packages (310, 810) by means of the forward projection module (300, 800), and performing a second pre-training (S330, S830) on the forward projection module (300, 800) with a second pre-training loss function determined based on visualisation information corresponding to the respective medical visualising data packages of the first auxiliary pre-training set (115, 615)of the medical visualising data packages and on the second pre-training medical visualising data packages (320, 820), and- generating, in a medical field-specific training step after the second pretraining step, a medical field-specific training medical visualising data package (420, 920) for each of those by processing a medical fieldspecific training set (410, 910) of volumetric visualising data packages generated based on a medical field-specific artificial investigation area model by means of the forward projection module (300, 800), generating medical field-specific training loss-check volumetric visualising data packages (430, 930) by processing the medical field-specific training medical visualising data packages (420, 920) by means of the reconstruction module (200, 700), and performing a first further training (S440, S940) and a second further training (S450, S950) on the reconstruction module (200, 700) and on the forward projection module (300, 800), respectively, with a model reproduction loss function determined based on the medical field-specific training set (410, 910) of the volumetric visualising data packages and on the medical field-specific training loss-check volumetric visualising data packages (430, 930).

2. The training method according to claim 1 , characterised by generating, in a post-training step after the medical field-specific training step, a post-training volumetric visualising data package (530, 1030) by processing a post-training input medical visualising data package (500, 1000) by means of the reconstruction module (200, 700), generating a post-training loss-check medical visualising data package (535, 1035) by processing the post-training volumetric visualising data package (530, 1030) by means of the forward projection module (300, 800), and performing a third further training (S540, S1040) and a fourth further training (S550, S1050) on the reconstruction module (200, 700) and on the forward projection module (300, 800), respectively, with a reproduction-appropriateness loss function determined based on the post-training input medical visualising data package (500, 1000) and the post-training loss-check medical visualising data package (535, 1035).

3. The training method according to claim 2, characterised in that the reproduction-appropriateness loss function is:- a Poisson-based maximum log-likelihood loss function, or- a structural similarity based loss function.

4. The training method according to any of claims 1 to 3, characterised in that in the medical field-specific training step, by composing the model reproduction loss function based on the volumetric visualising data packages of the medical field-specific training set (410, 910) generated based on the medical field-specific artificial investigation area model and on the medical field-specific training loss-check volumetric visualising data packages (430, 930)- first deviation data and second deviation data characterizing difference from healthy are calculated for the volumetric visualising data packages of the medical field-specific training set (410, 910) generated based on the medical field-specific artificial investigation area model, and for the medical field-specific training loss-check volumetric visualising data packages (430, 930), respectively, and- the model reproduction loss function is determined by a deviation between the first deviation data and the second deviation data.

5. The training method according to claim 4, characterised in that- the first deviation data is a first contrast-to-noise ratio, and the second deviation data is a second contrast-to-noise ratio, or- the first deviation data and the second deviation data are determined by a deviation evaluator module.

6. The training method according to any of claims 1 to 5, characterised in that quantum representation is applied in the system.

7. The training method according to any of claims 1 to 6, characterised in that such a system is applied which is adapted for modality-preserving reconstruction of a medical visualising data package, and the reconstruction module (200, 700) is configured as a modality-preserving reconstructionmodule (200) adapted for generating a modality-preserving volumetric visualising data package by processing a medical visualising data package.

8. The training method according to claim 7, characterised by- applying, in the first auxiliary pre-training set (115) of the medical visualising data packages, pre-training auxiliary medical visualising data package groups comprising medical visualising data packages with different noise levels each generated from a base set (110) of medical visualising data packages with a splitter module (105), and applying, in the second pre-training step, as the visualisation information corresponding to the respective given medical visualising data packages of the first auxiliary pre-training set (115) of the medical visualising data packages, from the base set (110) of the medical visualising data packages, the medical visualising data package corresponding to each of the medical visualising data packages of the first auxiliary pre-training set (115), and- the second auxiliary pre-training set (120) of the volumetric visualising data packages corresponding to the first auxiliary pre-training set (115) of the medical visualising data packages is generated by processing each of the medical visualising data packages of the base set (110) of the medical visualising data packages by means of a different-from-machine learning-based reconstruction module (100).

9. The training method according to claim 8, characterised by- for each of the medical visualising data packages of the base set (110) of the medical visualising data packages, a plurality of volumetric visualising data packages of a second auxiliary pre-training set (120) are generated by processing those, by means of a different-from-machine learningbased reconstruction module (100), with a plurality of parameter values being different from each other of reconstruction parameters (102) corresponding to the different-from-machine learning-based reconstruction module (100), and- the first pre-training loss function is composed based on a plurality of volumetric visualising data packages of the second auxiliary pre-trainingset (120) generated by processing with a plurality of parameter values different from each other of the reconstruction parameter (102) and on first pre-training volumetric visualising data packages (210) generated by processing the medical visualising data packages of the first auxiliary pretraining set (115) by means of the reconstruction module (200).

10. The training method according to any of claims 7 to 9, characterised in that a SPECT or PET modality is applied in the system, and that in addition to providing medical visualising data packages as an input to the reconstruction module (200) and volumetric visualising data packages as an input to the forward projection module (300), related CT anatomy information (125, 425, 525) or MRI anatomy information is also provided.11 . The training method according to any of claims 1 to 6, characterised in that such a system is applied which is adapted for a modality-switching reconstruction of a medical visualising data package, and the reconstruction module (200, 700) is configured as a modality-switching reconstruction module (700) adapted for generating a modality-switching volumetric visualising data package by processing a medical visualising data package.

12. The training method according to claim 11 , characterised by- applying, in the first auxiliary pre-training set (615) of the medical visualising data packages, pre-training auxiliary medical visualising data package groups comprising medical visualising data packages with different noise levels each generated from a base set (610) of medical visualising data packages with a splitter module (605), and applying, in the second pre-training step, as visualisation information corresponding to respective given medical visualising data packages of the first auxiliary pre-training set (615) of the medical visualising data packages, such medical visualising data packages of the base set (610) of the medical visualising data packages that correspond to each of the medical visualising data packages of the first auxiliary pre-training set (615) of the medical visualising data packages, and- the second auxiliary pre-training set (620) of the volumetric visualising data packages corresponding to the first auxiliary pre-training set (615) ofthe medical visualising data packages is generated by processing the modality-switching medical visualising data packages (625) corresponding to each of the medical visualising data packages of the base set (610) of the medical visualising data packages by means of a different-from-machine learning-based reconstruction module (600).

13. The training method according to claim 12, characterised by- for each of the medical visualising data packages of the base set (610) of the medical visualising data packages, a plurality of volumetric visualising data packages of a second auxiliary pre-training set (620) are generated by processing, by means of the different-from-machine learning-based reconstruction module (600), the modality-switching medical visualising data packages (625) with a plurality of parameter values different from each other of reconstruction parameters corresponding to the different- from-machine learning-based reconstruction module (600), and- the first pre-training loss function is composed based on a plurality of volumetric visualising data packages of the second auxiliary pre-training set (620) generated by processing with a plurality of parameter values different from each other of the reconstruction parameter and on first pretraining volumetric visualising data packages (710) generated by processing the medical visualising data packages of the first auxiliary pretraining set (615) by means of the reconstruction module (700).

14. The training method according to any of claims 1 to 13, characterised in that the visualising data package is a static image or a time-varying image.

15. A system for reconstructing a medical visualising data package, c h a r a c t e r i s e d by- comprising a reconstruction module (200, 700) based on machine learning, adapted for generating a volumetric visualising data package by processing a medical visualising data package, and- being trained by means of the training method according to any of claims 1 to 14.

16. The system according to claim 15, characterised by- further comprising a forward projection module (300, 800) based on machine learning, adapted for generating a medical visualising data package by processing a volumetric visualising data package, and- being trained applying the training method according to claim 2.

Images