Characterization of perfusion defects

Abstract

The invention relates to characterization of perfusion defects. In order to characterize a perfusion defect in a vascular structure of a blood supply of an organ (5, 8) of a patient on the basis of imaging, medical imaging data (11, 12, 17) are received, the medical imaging data (11, 12, 17) containing energy-resolved CT imaging data (11, 12, 17). A blood flow occlusion object (7) in the vascular structure is detected based on the medical imaging data (11, 12, 17). A perfusion defect score (16) of a target region (6a, 6b) of the at least one organ (5, 8) is determined from the energy-resolved CT imaging data (11, 12, 17), the blood perfusion of which is potentially influenced by the occluded object (7).

Description

TECHNICAL FIELD

[0001] The present invention relates to a computer-implemented method for characterizing perfusion defects in a vascular structure of at least one organ of a patient based on imaging, a corresponding data processing system and a corresponding computer program product. BACKGROUND

[0002] The present invention belongs to the field of medical imaging analysis for creating data-based, in particular visual, representations of the inside of a body based on detection of physical signals and imaging data-based analysis of resulting image data to support analysis, in particular visual representation, of the function of an organ or tissue.

[0003] Pulmonary embolism is a potentially life-threatening condition that can occur when a blood clot, also known as a thrombus, forms in the pulmonary arteries. The clot can block or partially obstruct blood flow to the lungs, leading to various symptoms and complications. In medical imaging data, such as computed tomography (CT) imaging data, a pulmonary embolism can for example manifest as a filling defect or a region of reduced contrast enhancement within a pulmonary artery.

[0004] It is known that algorithms based on machine learning models (MLM) can automatically detect blood clots in pulmonary arteries from medical imaging data. For example, the publication "Deep Learning-Based Algorithm for Automatic Detection of Pulmonary Embolism in Chest CT Angiograms" by P. A. Grenier et al., Diagn. Basel, 2023, vol. 13, no. 7, utilizes an artificial neural network (ANN) to detect such clots in CT angiography (CTA) imaging data.

[0005] However, detecting only the presence and location of blood clots does not necessarily enable a reliable assessment of the criticality of the resulting perfusion defects.

[0006] The publication "Segmental Anatomy of the Lungs. A Study of the Patterns of the Segmental Bronchi and Related Pulmonary Vessels" by E. A. Boyden, published in 1955 by Blakiston Division, McGraw-Hill Book Company, Inc., New York, describes a hierarchical classification of bronchi.

[0007] The publication "Quantification of lung perfusion blood volume (lung PBV) by dual-energy CT in patients with chronic thromboembolic pulmonary hypertension (CTEPH) before and after balloon pulmonary angioplasty (BPA): Preliminary results" by H. Koike et al., published in European Journal of Radiology, Volume 85, Issue 9, 2016, Pages 1607-1613 describes how lung perfusion blood volume PBV can be quantified by calculating the ratio of lung PBV to pulmonary artery enhancement in dual-energy CT (DECT) imaging data.

[0008] The U-Net architecture, which is a widely used convolutional neural network architecture for image segmentation, but can also be used for other computer vision tasks, especially image-to-image tasks, is described in "U-Net: Convolutional Networks for Biomedical Image Segmentation" by O. Ronneberger et al., arXiv: 1505.04597. SUMMARY

[0009] It is an object of the present application to provide a feasible approach for automatically characterizing perfusion defects in our vascular structure for blood supply to at least one organ of a patient.

[0010] This object is solved by the subject matter according to the present application. Further implementations and preferred embodiments are subject of the description.

[0011] The underlying idea of the present application is to not only detect blood flow obstruction objects in the vascular structure for blood supply to at least one organ of a patient, but also to determine a perfusion defect score for a target region potentially affected by the obstruction objects from energy-resolved CT imaging data.

[0012] According to an aspect of the present application, a computer-implemented method for characterizing perfusion defects in a vascular structure, in particular a blood vessel structure, of a blood supply of at least one organ of a patient based on imaging is provided. Therein, medical imaging data depicting the at least one organ and the vascular structure are received, e.g., from a medical imaging device or from a data storage device. The medical imaging data contain or consist of energy-resolved CT imaging data. Blood flow occlusion objects, e.g., blood clots, in the vascular structure are detected based on the medical imaging data. A perfusion defect score of a target region of the at least one organ is determined from the energy-resolved CT imaging data, wherein blood perfusion in the target region can potentially be affected by the blood flow occlusion objects.

[0013] Unless otherwise stated, all steps of the computer-implemented method can be performed by a data processing system comprising at least one data processing device. In particular, the at least one data processing device is configured or adapted for performing the steps of the computer-implemented method. To this end, the at least one data processing device may, for example, store a computer program containing instructions which, when executed by the at least one data processing device, cause the at least one data processing device to perform the computer-implemented method. The expressions "data processing system" and "at least one data processing device" can be used interchangeably herein and subsequently. The same applies to the respective expressions derived therefrom. In other words, a system according to the present application can be improved by the features described or claimed in the context of the method and vice versa. In this case, the functional features of the method are embodied by the target units or modules of the system.

[0014] In case the at least one data processing device comprises two or more data processing devices, certain steps performed by the at least one data processing device can also be understood such that different data processing devices perform different steps or different parts of a step. In particular, it is not required that each data processing device performs the steps completely. In other words, the execution of the steps can be distributed among the two or more data processing devices.

[0015] From each implementation of the computer-implemented method, by including the respective steps of generating the medical imaging data, e.g., by a medical imaging device, a respective implementation of the method for characterizing perfusion defects in a vascular structure of a blood supply of at least one organ of a patient based on imaging is obtained, which is not purely computer-implemented.

[0016] Energy-resolved CT imaging data can contain contrast-enhanced imaging data acquired by X-ray computed tomography (CT) using a contrast agent. The contrast agent for X-ray CT is typically an iodine-based contrast agent. This helps to highlight structures such as blood vessels which would otherwise be difficult to distinguish from the surrounding environment. The use of contrast material also helps to acquire functional information about the tissue. Typically, images are taken not only in the case of the use of a radiocontrast agent, but also in the case of the non-use of a radiocontrast agent.

[0017] Energy-resolved CT imaging data (also referred to as spectral-resolved CT imaging data) is so called because it contains imaging data for different energies of the detected X-rays or X-ray quanta, respectively. This can be achieved, for example, by using different X-ray spectra, for example by using different X-ray filters, by using X-ray sources for providing X-rays with different energy spectra, for example in a dual-source CT scanner, or by kV switching of a single X-ray source, or by using an energy-resolving X-ray detector, in particular by photon-counting CT (PCCT), wherein different energy thresholds can be used for counting incident photons. Medical imaging data, for example energy-resolved CT imaging data, can contain image data obtained from energy-resolved CT imaging data raw data, for example one or more reconstructed CT image volumes, for example including a conventional CT image volume, imaging data with material decomposition information, virtual mono- energy imaging data, iodine-enhanced CT image volumes, virtual non-contrast imaging data, dual energy ratio (DER), CT image volume DER, etc.

[0018] Energy-resolved CT imaging data allows to distinguish between different materials. This can be used, for example, by calculating a corresponding contrast agent map, for example an iodine map or a dual energy ratio, DER, map, which can then be analyzed to determine a perfusion defect score.

[0019] Detecting the occlusion object, in particular, contains determining a position of the occlusion object in the blood vessel structure or a region of the blood vessel structure, where the occlusion object is located. This information can be used, for example, to identify a target region. The occlusion object can also be detected from the energy-resolved CT imaging data. Alternatively, the medical imaging data can contain further image data and the occlusion object is detected from said further image data. For detecting the occlusion object, for example, known image analysis algorithms can be used.

[0020] If not stated otherwise, here and subsequently, CT imaging data can contain, for example, one or more three-dimensional CT reconstructions, also referred to as reconstructed CT image volumes.

[0021] The at least one organ may, for example, comprise a lung or a portion of a lung of a patient, in particular including the bronchial tree of the lung and / or the parenchymal tissue of the lung. The at least one organ may, for example, also include a vascular structure. The vascular structure may, for example, comprise a pulmonary artery structure or a pulmonary artery tree of the lung.

[0022] The perfusion defect score can be a final result of the computer-implemented method according to the present application. In particular, by determining the perfusion defect score, a perfusion defect in the vascular structure can be characterized. According to the perfusion defect score, the severity or criticality of the impact of the obstructive object may, for example, be assessed.

[0023] The size and location of the obstructive object can vary, from small to large and from peripheral to central. Thus, the impact of the obstructive object can be very small, but also severe and even life-threatening, as the obstructive object can reduce or completely cut off the blood supply to downstream lung tissue, resulting in a reduced lung perfusion. Thus, the perfusion defect score can be used to distinguish the severity of the obstructive object.

[0024] In particular, the perfusion defect score can be considered to be related or associated to a specific obstructive object. In other words, determining the perfusion defect score can be considered to classify the obstructive object according to the severity of its impact on the perfusion of the at least one organ.

[0025] In some implementations, the described steps of the computer-implemented method according to the present application can be performed twice or multiple times, respectively, for different obstructive objects at different locations. In this case, different obstructive objects can be ranked and / or filtered according to their perfusion defect scores.

[0026] According to several embodiments, a segmentation is generated based on the medical imaging data that divides the at least one organ into a plurality of segments, wherein the plurality of segments is hierarchically classified according to blood supply of the segments by the vascular structure. A target region is determined to comprise or consist of one or more target segments of the plurality of segments, whose blood perfusion is potentially affected by the obstructive object according to the hierarchical classification.

[0027] What is exploited in such embodiments is that the anatomical structure of a human organ, in particular a lung, is well known and thus also the relationship between segments of the organ and the respective blood vessels that supply them. Thus, by using a segmentation that includes a hierarchical classification, a particularly reliable characterization of the perfusion defect can be achieved.

[0028] The plurality of segments may, for example, be generated using known medical image segmentation algorithms.

[0029] In case at least one of the organs comprises a lung, the segmentation may, for example, be generated according to a bronchial lung segment model or according to a lung segmentation model, e.g. based on the hierarchical classification according to Boyden mentioned in the above section of the present disclosure. The hierarchical classification of the plurality of segments according to the blood supply by the vascular structure may, for example, be understood such that, for each segment of the plurality of segments, it is defined whether the blood supply of the respective segment by the vascular structure depends on the blood supply by the vascular structure of one or more other segments of the plurality of segments and, if applicable, on the blood supply by the vascular structure of which other segments of the plurality of segments. In other words, if a segment of the plurality of segments is given and it is assumed that the blood supply of the respective segment by the vascular structure is completely blocked due to an occlusion object in the respective portion of the vascular structure, it is defined which other segments of the plurality of segments are also cut off from the blood supply by the vascular structure as a consequence.

[0030] For example, a root segment of the plurality of segments can be identified by detecting the occlusion object, which is a segment of the plurality of segments, wherein the occlusion object is located in the respective portion of the vascular structure directly supplying the respective segment. In other words, the root segment is a segment of the plurality of segments, the blood supply of which is affected by the occlusion object, even though the blood supply of other segments, in particular of segments of a higher hierarchical level, is not affected by the occlusion object. The root segment can also be determined as the segment of the plurality of segments that is closest to the occlusion object.

[0031] The one or more target segments comprise the root segment. In some cases, the one or more target segments consist of the root segment. Alternatively, the one or more target segments comprise one or more downstream segments of the plurality of segments. In particular, the one or more target segments consist of the root segment and the one or more downstream segments. The one or more downstream segments can be segments, the blood supply of which via the vascular structure is cut off in case the blood supply of the root segment is cut off.

[0032] According to several embodiments, for each target segment of the one or more target segments, a respective segment perfusion defect score is determined based on the energy-resolved CT imaging data, and the perfusion defect score is determined from the segment perfusion defect scores, in particular from all segment perfusion defect scores of all target segments.

[0033] By individually assessing the impact of the occlusion object for each target segment, a comprehensive analysis of the overall perfusion defect can be made and, thus, a particularly reliable characterization of the perfusion defect.

[0034] For example, the perfusion defect score can be calculated as a sum of the segmental perfusion defect scores of the target segments or as an average of the segmental perfusion defect scores of the target segments or as a maximum segmental perfusion defect score of the target segments, etc. The perfusion defect score can also be calculated as a number of target segments whose segmental perfusion defect score is at least a predefined threshold value.

[0035] According to several embodiments, a perfusion blood volume PBV can be determined for each target segment from the energy-resolved CT imaging data and a segmental perfusion defect score can be calculated from the respective PBV.

[0036] The PBV can reliably quantify the blood perfusion in the target segment and thus the influence of the obstructive object on the blood perfusion in the target segment. In particular, the PBV can be determined based on the energy-resolved CT imaging data by quantifying the iodine uptake of the contrast agent, e.g. iodine, by the organ, in particular the lung parenchyma. For example, the lower the PBV and / or the larger the region of reduced PBV, the greater the influence of the obstructive object on the blood perfusion of the respective target segment and thus the greater the respective segmental perfusion defect score.

[0037] According to several embodiments, for each of the one or more target segments, a size of a perfusion defect region in the respective target segment, e.g. a volume of the perfusion defect region, is determined based on the energy-resolved CT imaging data. The respective segmental perfusion defect score is determined from or as the size of the perfusion defect region.

[0038] The size of the perfusion defect region can for example be determined as a volume in the respective target segment, wherein the PBV is for example smaller than a predefined threshold value.

[0039] The size of the perfusion defect region is a reliable measure to assess the influence of the obstructive object on the blood perfusion of the respective target segment. Thus, the segmental perfusion defect score and thus the perfusion defect score can be determined in a particularly reliable manner.

[0040] For example, the perfusion defect score can be calculated as a total perfusion defect size, or in other words, a total perfusion defect volume, which is given by a sum of the sizes of the perfusion defect regions of the individual target segments. The perfusion defect score can also be calculated as a ratio of the total perfusion defect size to a predetermined or predefined total lung volume of the patient.

[0041] According to several embodiments, at least one PBV value of the target region is determined from the energy-resolved CT imaging data and the perfusion defect score is determined from the at least one PBV value.

[0042] The at least one PBV value can for example be a single PBV value for the entire target region or a respective segmental PBV value for each target segment.

[0043] PBV values ​​can be derived from the PBV of the target region or the corresponding target segment. For example, at least one PBV value may correspond to the total perfusion defect size, or at least one PBV value may contain the corresponding segment PBV value for each of the one or more target segments, such as the corresponding size of the perfusion defect region of the corresponding target segment.

[0044] According to several embodiments, medical imaging data, particularly energy-resolved CT imaging data, includes photon-counted CT (PCCT) image data and detects obstructed objects based on the PCCT image data.

[0045] Such an implementation is particularly advantageous because, on the one hand, PCCT allows for ultra-high resolution CT reconstruction, and on the other hand, data in the form of energy resolution can naturally be provided by evaluating different energy thresholds while counting incident X-ray photons. Therefore, obstructed objects can be detected with particularly high accuracy, and perfusion defect fractions can be obtained with particularly reliable accuracy.

[0046] A photon-counting detector for computed tomography (CT) generates spectrally resolved computed tomography medical image data. The photon-counting detector can be configured to acquire X-ray projection data from multiple energy chambers. These window parameters of the energy chambers are defined by energy thresholds corresponding to the (spectral) energies of the X-ray photons to be detected. The detector's energy thresholds can be predetermined and set by corresponding control logic implemented within the photon-counting detector. Multiple (e.g., four) energy chambers can be used to cover the energy range of the acquired X-ray projection data.

[0047] The window parameters of the energy chamber can be predefined before acquiring raw energy-resolved imaging data. To improve the quality of the acquired photon-counting spectral computed tomography (CT) data, the window parameters of the energy chamber can be adjusted according to the specific CT application. For example, the window parameters of the energy chamber can be pre-selected to obtain energy-resolved imaging data that allows for optimized material decomposition in energy-resolved imaging, such as any material selected from the group consisting of iodine, calcium, and water. In particular, providing three or more energy chambers to achieve decomposition of three materials may be advantageous. Thus, energy-resolved image data optimized for tasks such as detecting perfusion defects or vascular occlusion can be acquired.

[0048] In an embodiment of segmentation generation based on energy-resolved CT imaging data, segmentation may be generated, for example, based on PCCT imaging data.

[0049] According to several embodiments, segmentation is generated by applying a trained first machine learning model (MLM) to first input data that at least partially contains medical imaging data, such as energy-resolved CT imaging data.

[0050] Generally, a trained MLM can mimic cognitive functions of a human being related to thinking of other human beings. In particular, by training based on training data, the MLM can be able to adapt to new situations and detect and infer patterns. Another term for a trained MLM is a “trained function”.

[0051] For example, the trained MLM can be trained using energy-resolved image data as training data, wherein the training data has been annotated with additional training information. The additional training information can comprise annotations of ground truth information. As a non-limiting list of examples, the ground truth information can comprise classification information classifying a target region of the image data into one of a plurality of classes, scoring information associating a target region of the image with a score, possibly a disease score, image segmentation information, perfusion information, an indication of a presence or absence of a vessel obstruction in a target region of the image data, or an indication of a presence or absence of a perfusion defect in a target region. The annotations can be performed manually by trained personnel and / or automatically by using or with the aid of respective annotation software.

[0052] Generally, parameters of the MLM can be adjusted or updated by training. In particular, supervised training, semi-supervised training, unsupervised training, reinforcement learning, and / or active learning can be used. Further, representation learning, also referred to as feature learning, can be used. In particular, the parameters of the MLM can be adjusted iteratively by several training steps. In particular, within the training, a certain loss function, also referred to as cost function, can be minimized. In particular, in the training of artificial neural networks, ANN, a backpropagation algorithm can be used.

[0053] In particular, the MLM can comprise an ANN, a support vector machine, a decision tree, and / or a Bayesian network, and / or the MLM can be based on k-means clustering, Q-learning, a genetic algorithm, and / or association rules. In particular, the ANN can be or comprise a deep neural network, a convolutional neural network, or a convolutional deep neural network. Further, the ANN can be an adversarial network, a deep adversarial network, and / or a generative adversarial network (GAN).

[0054] The first MLM can be a known MLM for medical image segmentation, e.g., a U-Net based ANN, which has been trained in a conventional way.

[0055] According to several embodiments, the occluded object is detected by applying the trained second MLM to second input data at least partially comprising medical imaging data, e.g., energy-resolved CT imaging data.

[0056] The second MLM can be a known MLM for medical image segmentation or object detection in medical images, e.g. a U-Net based ANN or an algorithm as proposed by Grenier et al. in the above publication, which has been trained in a conventional manner.

[0057] In some embodiments, the first and second MLMs can also be part of a further MLM. In such embodiments, the obstructing object is detected and the segmentation is generated by applying the trained further MLM to at least a portion of the medical imaging data, e.g. the energy-resolved CT imaging data.

[0058] According to several embodiments, the energy-resolved CT imaging data comprises contrast-enhanced CT (CECT) imaging data, which e.g. includes a contrast agent map, in particular an iodine map, and / or a DER map.

[0059] Thus, in such embodiments, the perfusion defect score, in particular the segment perfusion defect score and / or the size of the perfusion defect region and / or the PBV value can be determined with increased accuracy.

[0060] According to several embodiments, the perfusion defect score is determined by applying a trained third MLM to third input data comprising the medical imaging data.

[0061] The third MLM can e.g. be based on a known MLM architecture, e.g. a U-Net trained in a conventional manner. In particular, the training of the third MLM can be based on a plurality of training data sets, each training data set comprising respective training imaging data corresponding to the medical imaging data and a corresponding perfusion defect score as a ground truth label.

[0062] Thus, the trained third MLM can directly predict the perfusion defect score. In such embodiments, the computer-implemented method according to the present application is particularly efficient from a computational perspective.

[0063] For example, as a result of applying the third MLM to the third input data, the output of the third MLM can also comprise the location of the obstructing object. In this case, the training data sets can also comprise the corresponding location as a ground truth label.

[0064] According to several embodiments, the third input data comprises a segmentation comprising a hierarchical classification.

[0065] In the respective training phase, the training data sets also comprise a corresponding training segmentation as input to the third MLM. In such embodiments, the training effort of the third MLM can be reduced, as the third MLM is given additional information about the content of the training imaging data. Furthermore, the accuracy of the output of the trained third MLM can also be improved.

[0066] According to another aspect of the present application, a data processing system is provided. The data processing system is configured to perform the computer-implemented method according to the present application.

[0067] In the present disclosure, the expressions “data processing system” and “at least one data processing device” can be used interchangeably. A data processing device can especially be understood as a data processing device comprising a processing circuitry. Thus, a data processing device can especially process data to perform a computational operation. This can also include operations performing an indexed access to a data structure, e.g. a lookup table (LUT), as well as data processing procedures implemented in hardware.

[0068] Especially, a data processing device can comprise one or more computers, one or more microcontrollers, and / or one or more integrated circuits, e.g. one or more application-specific integrated circuits ASIC, one or more field-programmable gate arrays FPGA, and / or one or more system-on-chips SoC. A data processing device can also comprise one or more processors, e.g. one or more microprocessors, one or more central processing units CPU, one or more graphics processing units GPU, and / or one or more signal processors, especially one or more digital signal processors DSP. A data processing device can further comprise a physical or virtual cluster of computers or other units.

[0069] In various embodiments, a data processing device comprises one or more hardware and / or software interfaces and / or one or more storage units.

[0070] A storage unit can be implemented as a volatile data storage, e.g. a dynamic random access memory DRAM or a static random access memory SRAM, or as a non-volatile data storage, e.g. a read-only memory ROM, a programmable read-only memory PROM, an erasable programmable read-only memory EPROM, an electrically erasable programmable read-only memory EEPROM, a flash memory or flash EEPROM, a ferroelectric random access memory FRAM, a magnetoresistive random access memory MRAM, or a phase change random access memory PCRAM.

[0071] Especially, a data processing system can comprise

[0072] an input data interface configured to receive medical imaging data depicting a vascular structure in at least one organ, and the vascular structure is received, wherein the medical imaging data comprises energy-resolved CT imaging data,

[0073] a detection module for detecting a blood flow obstruction object in the vascular structure based on the medical imaging data; and

[0074] - Determination module for determining the perfusion defect fraction of a target region of at least one organ, the blood perfusion of which is potentially affected by the obstructed object, wherein determination is performed based on energy-resolved CT imaging data.

[0075] According to another aspect of the present invention, a medical imaging system is provided. This medical imaging system includes a data processing system according to the present invention and a CT apparatus configured to generate medical imaging data depicting the vascular structures of at least one organ.

[0076] In particular, the CT device is configured to generate energy-resolved CT imaging data. For example, the CT device may be a photon-counting CT device.

[0077] Further implementations of the medical imaging system according to the invention directly follow various embodiments of the computer-implemented method according to the invention, and vice versa. In particular, the individual features, corresponding explanations, and advantages relating to the various implementations of the computer-implemented method according to the invention can be similarly applied to corresponding implementations of the medical imaging system according to the invention. In particular, the medical imaging system according to the invention is designed or programmed to perform the computer-implemented method according to the invention. In particular, the medical imaging system according to the invention performs the computer-implemented method according to the invention.

[0078] According to another aspect of the present invention, a computer program comprising instructions is provided. When a data processing system executes these instructions, the instructions cause the data processing system to perform a computer-implemented method according to the present invention.

[0079] For example, the instructions can be provided as program code. The program code can be provided, for example, as binary code or source code of an assembler and / or a programming language (such as C) and / or a program script (such as Python).

[0080] According to another aspect of the present invention, a computer-readable storage medium for storing a computer program according to the present invention is provided.

[0081] A computer program and a computer-readable storage medium are computer program products containing instructions, respectively.

[0082] Further features and combinations of features of the present invention can be obtained from the accompanying drawings, their description, and the claims. In particular, alternative implementations of the present invention may not necessarily include all the features of one of the claims. Alternative implementations of the present invention may include features or combinations of features not recited in the claims.

[0083] The present invention will now be explained in detail with reference to specific exemplary embodiments and corresponding schematic diagrams. In the accompanying drawings, elements that are the same or have the same function can be represented by the same reference numerals. The description of elements that are the same or have the same function need not be repeated in different drawings. Attached Figure Description

[0084] In the attached diagram,

[0085] Figure 1 An exemplary embodiment of the medical imaging system according to the present invention is shown;

[0086] Figure 2 A schematic flowchart illustrating an exemplary embodiment of a computer-implemented method according to the present invention for perfusion defects in vascular structures representing the blood supply to at least one organ of a patient based on imaging.

[0087] Figure 3 The illustration schematically shows a bronchial tree and obstruction object in an example of another exemplary embodiment of the method for computer-implemented perfusion defect characterization based on imaging, according to the present invention;

[0088] Figure 4 The illustration schematically shows a bronchial tree and obstruction object in an example of another exemplary embodiment of the method for computer-implemented perfusion defect characterization based on imaging, according to the present invention;

[0089] Figure 5 The illustration schematically shows a segmented cross-sectional view of a lung in an example of another exemplary embodiment of the method for computer-implemented perfusion defect characterization based on imaging, according to the present invention.

[0090] Figure 6 The illustration schematically shows a segmented sagittal cross-sectional view of a lung in an example of another exemplary embodiment of the method for computer-implemented perfusion defect characterization based on imaging, according to the present invention.

[0091] Figure 7 A schematic flowchart illustrating another exemplary embodiment of the method for computer implementation based on imaging characterization according to the present invention;

[0092] Figure 8 A schematic flowchart illustrating another exemplary embodiment of the method for computer implementation based on imaging characterization according to the present invention;

[0093] Figure 9 An exemplary embodiment of an artificial neural network is illustrated schematically;

[0094] Figure 10 An exemplary embodiment of a convolutional neural network is illustrated schematically; and

[0095] Figure 11 Another exemplary embodiment of a convolutional neural network is schematically shown. DETAILED DESCRIPTION

[0096] Figure 1 An exemplary embodiment of a medical imaging system 1 according to the present application is shown. The medical imaging system 1 comprises a data processing system 4 according to the present application and a CT device 2, which is implemented as a photon counting CT device, among others.

[0097] The CT device 2 is configured to generate energy-resolved CT imaging data 17, which comprises, among others, one or more reconstructed CT image volumes depicting at least one organ 5, 8 of a patient and a vascular structure for a blood supply of the at least one organ 5, 8. The at least one organ 5, 8 is given by the patient’s lung 8, including the bronchial tree 5 (as schematically shown in Figure 3 and Figure 4 ) and the parenchymal tissue of the lung 8 (as schematically shown in Figure 5 and Figure 6 ). The vascular structure may, for example, include or correspond to the pulmonary arteries.

[0098] The data processing system 4 is configured to perform an exemplary embodiment of a computer-implemented method for imaging-based characterization of perfusion defects in a vascular structure according to the present application. Figure 2 、 Figure 7 and Figure 8 schematically flowcharts of different exemplary embodiments of the computer-implemented method are shown.

[0099] In general, for implementing the computer-implemented method according to the present application, the data processing system 4 receives the energy-resolved CT imaging data 17 and detects a blood flow obstruction object 7, in particular a blood clot, in the vascular structure based on the energy-resolved CT imaging data 17. The data processing system 4 determines a perfusion defect score 16 for a target region 6a, 6b of the at least one organ 5, 8, wherein the blood perfusion of the target region 6a, 6b is potentially affected by the obstruction object 7 according to the energy-resolved CT imaging data 17.

[0100] According to the embodiment of the computer-implemented method as depicted in Figure 2 , the obstruction object 7 is detected in step 210, e.g., using a known algorithm (e.g., an algorithm based on a trained first MLM) for clot detection in the pulmonary arteries. In addition to the known method, the detection of the obstruction object 7 can be improved, e.g., by using spectral information of the energy-resolved CT imaging data 17 (e.g., material labels) to better distinguish between contrast agent (in particular iodine) and clot material.

[0101] In step 220, a segmentation 13 is generated based on the medical imaging data 17 that divides the at least one organ 5, 8 into a plurality of segments 6, 9, wherein the plurality of segments 6, 9 is hierarchically classified according to blood supply of these segments by vascular structures.

[0102] The segmentation may, for example, be a segmentation of the bronchial tree 5 as shown in Figure 3 and Figure 4 The segmentation may, for example, be based on a classification according to Boyden. The trachea branches at the carina into two main bronchi, which are also referred to as primary bronchi. One main bronchus leads to the right lung 8b and the other main bronchus leads to the left lung 8a. Each main bronchus enters the respective lung and further branches. The main bronchi branch into lobar bronchi, which are also referred to as secondary bronchi. The right lung 8b has three lobes, thus, the right main bronchus branches into three lobar bronchi, which are referred to as upper, middle and lower lobar bronchi. The left lung 8a has two lobes, thus, the left main bronchus branches into two lobar bronchi, namely the upper and lower lobar bronchi. The lobar bronchi further branch into segmental bronchi. The segmental bronchi supply a specific segment of lung tissue, known as bronchopulmonary segments. Each lung has a different number of bronchopulmonary segments, wherein the right lung 8b has more bronchopulmonary segments than the left lung 8a. The segmental bronchi continue to branch into smaller bronchi, which are known as subsegmental bronchi. These subsegmental bronchi supply smaller areas within the bronchopulmonary segments.

[0103] The bronchial tree 5 is not only responsible for the flow of air, but also requires blood supply of the pulmonary arteries themselves. The pulmonary arteries branch off from the systemic circulation and supply oxygenated blood to the bronchi, bronchioles and other lung structures. Each segment 6 of the bronchial tree 5 has its own pulmonary artery branch, and thus, as shown in Figure 5 and Figure 6 each pulmonary segment 9 is a part of the lung that is supplied by its own bronchi and arteries. Each segment is functionally and anatomically discrete, allowing for surgical resection of an individual segment without affecting its neighboring segments. Thus, each segment 6 of the bronchial tree 5 can also be associated with a respective segment 9 of the parenchymal tissue of the lung 8, and in particular, an occlusion of blood flow in a part of the vascular structure that supplies a certain segment 6 of the bronchial tree 5 potentially affects the related blood perfusion of the corresponding segment 9 of the parenchymal tissue of the lung 8.

[0104] The bronchial classification according to Boyden refers to a standard nomenclature for describing the anatomical structure of bronchopulmonary segments. The described model of bronchopulmonary segments is a preferred segmentation. However, the segmentation can also be based on another segment model, e.g. lobe, left lung and right lung 8b, or another segment model definition of custom-made lung segments.

[0105] In Figure 2In the method of Fig. 2, the target regions 6a, 6b are determined to comprise one or more target segments 6a, 6b of the plurality of segments 6, 9, whose blood perfusion is potentially affected by the obstructive object 7 according to the hierarchical classification. Notably, for illustrative reasons, Figure 3 and Figure 4 The position of the obstructive object 7 in certain segments 6a of the bronchial tree 5 is shown. This does not mean that the object 7 is located within the respective bronchus, but in the corresponding part of the vascular structure.

[0106] As shown in Figure 5 and Figure 6 The parenchymal tissue of the lung 8 can be segmented based on known algorithms, e.g. based on a trained second MLM. For example, the second MLM can be trained on ultra-high resolution photon counting CT data, which better depict the inter-segment fissures, allowing a better anatomical depiction of the lung segments.

[0107] In step 230, the root segment 6a of the bronchial segments 6, which is the segment 6 corresponding to the position of the obstructive object 7, and / or the corresponding root segment of the parenchymal tissue of the lung 8 are automatically identified based on the segmentation 13.

[0108] Based on the hierarchical anatomical model of the bronchial lung segments, all dependent downstream segments 6b of the bronchial tree 5 and / or all dependent downstream segments of the parenchymal tissue of the lung 8 are identified in step 240.

[0109] In each downstream segment of the parenchymal tissue of the lung 8, the parenchymal perfusion is quantified in step 250 based on spectral perfusion data acquired from the energy-resolved CT imaging data 17, e.g. based on iodine absorption or dual energy ratio DER. For example, the spectral perfusion data can be normalized based on reference values in a reference region, e.g. the pulmonary trunk, to account for contrast bolus variations between different scans and patients. This quantification can be performed, for example, by comparing the iodine concentration, enhancement value or DER between the downstream segments of the parenchymal tissue of the lung 8 and normal lung segments. For example, normal lung segments can be defined as the perfusion in a normal reference population or the perfusion in unaffected downstream segments in a given patient.

[0110] In step 260, a perfusion defect score 16 is then summarized from the segment-by-segment perfusion defect quantification. The perfusion defect score 16 can be determined, for example, as the number of segments with perfusion defects or significant perfusion defects, the total volume of perfusion defect regions, the percentage of lung involvement calculated by dividing the volume of perfusion defects by the total lung volume, etc.

[0111] Steps 210 to 260 can also be repeatedly performed for different relevant obstructive objects 7, e.g. see Figure 3 and Figure 4comparisons. In particular, the associated obstructive objects 7 can be ranked or filtered according to the severity of the obstructive objects as given by the perfusion defect scores 16 of the obstructive objects. Furthermore, in some embodiments, perfusion defects that cannot be explained by the associated obstructive objects 7 can be marked for further analysis.

[0112] Figure 7 A schematic flow chart illustrating another exemplary embodiment of the computer-implemented method for imaging-based characterization according to the present application is shown. Therein, the perfusion defect scores 16 are determined by applying a trained third MLM 10 to third input data comprising energy-resolved CT imaging data 17. In particular, in this embodiment, the third MLM 10 is trained on the auxiliary task of predicting the segmentation 13 as part of its output. Figure 7 In embodiments of the above-described

[0113] The third MLM 10 receives as input, for example, the CECT image volume 11 (preferably with ultra-high resolution) and the spectral information 12 (e.g. the associated iodine image or DER image). Optionally, the segmentation 13 with hierarchical labels can be passed as additional input to the third MLM 10 to guide the third MLM 10 to respect given lung segments. If lung segments are not passed as input, the third MLM 10 may, for example, identify perfusion defect regions that do not conform to any prior segment model.

[0114] The output of the third MLM 10 comprises, for example, the respective locations of the detected obstructive objects 7. Furthermore, the detection algorithm can also be trained on the spectral information 12, which, for example, helps to distinguish between blood clot material and iodine. The output of the third MLM 10 further comprises the respective perfusion defect scores 16. Optionally, the output of the third MLM 10 comprises the respective detected regions with perfusion defects and the corresponding perfusion defect quantification per obstructive object.

[0115] Figure 8 A schematic flow chart illustrating another exemplary embodiment of the computer-implemented method for imaging-based characterization according to the present application is shown. As Figure 7 described, the perfusion defect scores 16 are determined by applying a trained third MLM 10 to third input data comprising energy-resolved CT imaging data 17. The input of the third MLM 10 comprises, for example, the CECT image volume 11 and the spectral information 12, but not the segmentation 13. Instead, in this embodiment, the third MLM is trained on the auxiliary task of predicting the segmentation 13 as part of its output.

[0116] For example, the above-described Figure 7 and Figure 8The third MLM 10 of embodiments is trained in a supervised manner based on expert annotations of the output data to be estimated.

[0117] In several embodiments of the computer-implemented method according to the present application, clot detection and assessment of parenchymal perfusion defects are combined to improve detection of severe pulmonary embolism thrombi with associated perfusion defects.

[0118] In several embodiments, PCCT imaging data is used as a basis for the inherent advantages of the ultra-high resolution and spectral image information provided with PCCT scans. This helps to distinguish between pulmonary emboli that have no serious impact on the patient and more serious pulmonary emboli. For example, blood clots can have a high density appearance that is difficult to distinguish on contrast-enhanced conventional CT images. However, blood clots and iodine have different material compositions, which can be more accurately assessed by spectral PCCT.

[0119] As mentioned above, several method steps can be performed using respectively trained MLMs, e.g. ANNs. Figure 9 An embodiment of an ANN 800 is shown, which is designed as a MLP, for example. The ANN 800 comprises nodes 820,..., 832 and edges 840,..., 842, wherein each edge 840,..., 842 is a directed connection from a first node 820,..., 832 to a second node 820,..., 832. In general, the first node 820,..., 832 and the second node 820,..., 832 are different nodes 820,..., 832. However, the first node 820,..., 832 and the second node 820,..., 832 can also be identical. For example, in the ANN 800 shown in Fig. 8, the edge 840 is a directed connection from the node 820 to the node 823 and the edge 842 is a directed connection from the node 830 to the node 832. The edges 840,..., 842 from the first node 820,..., 832 to the second node 820,..., 832 are also denoted as incoming edges of the second node 820,..., 832 and outgoing edges of the first node 820,..., 832. Figure 9 In the ANN 800 shown in Fig. 8, the edge 840 is a directed connection from the node 820 to the node 823 and the edge 842 is a directed connection from the node 830 to the node 832. The edges 840,..., 842 from the first node 820,..., 832 to the second node 820,..., 832 are also denoted as incoming edges of the second node 820,..., 832 and outgoing edges of the first node 820,..., 832.

[0120] In the present example, the nodes 820,..., 832 of the artificial neural network 800 can be arranged in layers 810,..., 813, wherein these layers can contain an inherent order introduced by edges 840,..., 842 between the nodes 820,..., 832. In particular, edges 840,..., 842 can only exist between adjacent layers of nodes. In the shown example, there is an input layer 810 containing only nodes 820,..., 822 without incoming edges, an output layer 813 containing only nodes 831, 832 without outgoing edges, and hidden layers 811, 812 between the input layer 810 and the output layer 813. In general, the number of hidden layers 811, 812 can be chosen arbitrarily. In MLPs, this number is at least one. The number of nodes 820,..., 822 within the input layer 810 is typically related to the number of input values of the artificial neural network 800, and the number of nodes 831, 832 within the output layer 813 is typically related to the number of output values of the artificial neural network 800.

[0121] In particular, real numbers can be assigned as values to each node 820,..., 832 of the artificial neural network 800. Here, x(n) i denotes the value of the i-th node 820,..., 832 of the n-th layer 810,..., 813. The values of the nodes 820,..., 822 of the input layer 810 correspond to the input values of the artificial neural network 800. The values of the nodes 831, 832 of the output layer 813 correspond to the output values of the artificial neural network 800. Furthermore, each edge 840,..., 842 can contain a real number weight. In particular, the weights are real numbers within the interval [-1, 1] or within the interval [0, 1]. Here, w (m,n) i,j denotes the weight of the edge between the i-th node 820,..., 832 of the m-th layer 810,..., 813 and the j-th node 820,..., 832 of the n-th layer 810,..., 813. Furthermore, the weights w (n,n+1) i,j The abbreviation w (n) i,j is defined. In particular, for calculating the output values of the neural network 800, the input values are propagated through the neural network 800. In particular, the values of the nodes 820,..., 832 of the (n+1)-th layer 810,..., 813 can be calculated based on the values of the nodes 820,..., 832 of the n-th layer 810,..., 813 by:

[0122]

[0123] Here, the function f is denoted as a transfer function or activation function. Known transfer functions are step functions, sigmoid functions, such as logistic functions, generalized logistic functions, hyperbolic tangent functions, arctangent functions, error functions, smoothed step functions, or rectified functions. For example, transfer functions are used for normalization purposes. In particular, these values ​​are propagated layer by layer through the neural network 800, where the values ​​of the input layer 810 are given by the input of the neural network 800, where the values ​​of the first hidden layer 811 can be calculated based on the values ​​of the input layer 810 of the neural network 800, where the values ​​of the second hidden layer 812 can be calculated based on the values ​​of the first hidden layer 811, and so on.

[0124] To set the edge value w (m,n) i,j Training data is required to train the neural network 800. Specifically, the training data includes training input data and training output data (denoted as t). i For the training step, the neural network 800 is applied to the training input data to generate computed output data. Specifically, both the training data and the computed output data contain a number of values ​​equal to the number of nodes in the output layer. In particular, the comparison between the computed output data and the training data is used to recursively adjust the weights within the neural network 800 (backpropagation algorithm). Specifically, the weights are changed according to the following:

[0125]

[0126] Where γ is a predefined learning rate, and if the (n+1)th layer is not an output layer, then the numerical δ is... (n) j It can be based on δ (n+1) Recursively calculated as:

[0127]

[0128] And if the (n+1)th layer is the output layer 813, then the number δ (n) j Calculated as:

[0129]

[0130] Where f' is the first derivative of the activation function, and t (n+1) j It is the comparison training value of the j-th node of the output layer 813.

[0131] For example, the ANN can be a convolutional neural network, CNN. A CNN is a kind of ANN that uses convolution operations instead of general matrix multiplication in at least one of its layers. These layers are called convolutional layers. In particular, a convolutional layer performs a dot product of one or more convolutional kernels with the input data of the convolutional layer, wherein the entries of the one or more convolutional kernels are parameters or weights that can be adjusted by training. In particular, a Frobenius inner product and a ReLU activation function can be used. A convolutional neural network can contain additional layers, such as pooling layers, fully connected layers and / or normalization layers.

[0132] By using a convolutional neural network, the input can be processed very efficiently, because the convolution operations based on different kernels can extract various image features, so that the relevant image features can be found during training by adjusting the weights of the convolutional kernels. Furthermore, based on the weight sharing in the convolutional kernels, fewer parameters need to be trained, which can prevent overfitting in the training phase and allow for faster training or more layers in the network, thereby improving the performance of the network.

[0133] Figure 10 An exemplary embodiment of a convolutional neural network 700 is shown. In the shown embodiment, the convolutional neural network 700 contains an input node layer 710, a convolutional layer 711, a pooling layer 713, a fully connected layer 714 and an output node layer 716, as well as hidden node layers 712, 714. Alternatively, the convolutional neural network 200 can contain multiple convolutional layers 711, multiple pooling layers 713 and / or multiple fully connected layers 715, as well as other types of layers. The order of the layers can be chosen arbitrarily, typically using a fully connected layer 715 as the last layer before the output layer 716.

[0134] In particular, within the convolutional neural network 700, the nodes 720, 722, 724 of the node layers 710, 712, 714 can be considered to be arranged as a d-dimensional matrix or d-dimensional image. In particular, in the two-dimensional case, the value of a node 720, 722, 724 indexed by i and j in the nth node layer 710, 712, 714 can be denoted as x(n)[i,j]. However, the arrangement of the nodes 720, 722, 724 of one node layer 710, 712, 714 does not have an impact on the calculations performed within the convolutional neural network 700, because these calculations are only given by the structure and the weights of the edges.

[0135] The convolutional layer 711 is a connection layer between a preceding node layer 710 with node values x(n-1) and a following node layer 712 with node values x(n). In particular, the convolutional layer 711 is characterized by a structure and weights of the incoming edges of the convolutional operation based on a certain number of kernels. In particular, the structure and weights of the edges of the convolutional layer 711 are chosen such that the value x(n) of a node 722 of the following node layer 712 is computed based on the value x(n-1) of a node 720 of the preceding node layer 710 as a convolution x(n) = K * x(n-1), where the convolution * is defined in the two-dimensional case as:

[0136]

[0137] Here, the kernel K is a d-dimensional matrix, in the current example a two-dimensional matrix, which is typically small compared to the number of nodes 720, 722, e.g. a 3x3 matrix or a 5x5 matrix. In particular, this means that the weights of the edges in the convolutional layer 711 are not independent, but chosen to yield the convolution equation. In particular, for a kernel of a 3x3 matrix, there are only 9 independent weights, each entry of the kernel matrix corresponding to one independent weight, independent of the number of nodes 720, 722 in the preceding node layer 710 and the following node layer 712.

[0138] In general, the convolutional neural network 700 uses node layers 710, 712, 714 with multiple channels, in particular due to the use of multiple kernels in the convolutional layer 711. In these cases, the node layers can be seen as (d+1)-dimensional matrices, the first dimension indexing these channels. The action of the convolutional layer 711 is defined in the two-dimensional example as:

[0139]

[0140] where, corresponding to the a-th channel of the preceding node layer 710, corresponding to the b-th channel of the following node layer 712, and K a,b corresponding to one of the kernels. If the convolutional layer 711 acts on a preceding node layer 710 with A channels and outputs a following node layer 712 with B channels, there are A-B independent d-dimensional kernels K a,b .

[0141] In general, activation functions can be used in convolutional neural networks 700. In the present embodiment, ReLU (Rectified Linear Unit) is employed, where R(z) = max(0,z), so that the action of the convolutional layer 711 in the two-dimensional example is:

[0142]

[0143] It is also possible to use other activation functions, such as ELU (exponential linear unit), LeakyReLU, Sigmoid, Tanh, or Softmax.

[0144] In the illustrated embodiment, the input layer 710 comprises 36 nodes 720 arranged as a two-dimensional 6x6 matrix. The first hidden node layer 712 comprises 72 nodes 722 arranged as two two-dimensional 6x6 matrices, each of the two matrices being the result of a convolution of the values of the input layer with a 3x3 kernel within the convolution layer 711. Equivalently, the nodes 722 of the first hidden node layer 712 can be interpreted as being arranged as a three-dimensional 2x6x6 matrix, wherein the first dimension corresponds to the channel dimension.

[0145] An advantage of using the convolution layer 711 is that the spatial local correlation of the input data can be exploited by enforcing a local connection pattern between the nodes of adjacent layers, in particular by each node being connected to only a small area of nodes of the previous layer.

[0146] The pooling layer 713 is a layer of connections between a preceding node layer 712 having node values x(n-1) and a following node layer 714 having node values x(n). In particular, the pooling layer 713 can be characterized by a structure of edges and weights and an activation function forming a pooling operation based on a non-linear pooling function f. For example, in the two-dimensional case, the values x(n) of the nodes 724 of the following node layer 714 can be computed based on the values x(n-1) of the nodes 722 of the preceding node layer 712 as:

[0147]

[0148] In other words, by using the pooling layer 713, the number of nodes 722, 724 can be reduced by replacing a number d1-d2 of adjacent nodes 722 in the preceding node layer 712 by a single node 722 in the following node layer 714 (computed from the values of the number of adjacent nodes). In particular, the pooling function f can be a max function, an average, or an L2 norm. In particular, for the pooling layer 713, the weights of the edges are fixed and not modified by training.

[0149] An advantage of using the pooling layer 713 is that the number of nodes 722, 724 and the number of parameters is reduced. This results in a reduction of the amount of computation in the network and controlling overfitting.

[0150] In the illustrated embodiment, the pooling layer 713 is a max pooling layer, which replaces four adjacent nodes by only one node, the value being the maximum of the values of the four adjacent nodes. Max pooling is applied to each d-dimensional matrix of the preceding layer. In the present embodiment, max pooling is applied to each of the two two-dimensional matrices, reducing the number of nodes from 72 to 18.

[0151] Generally, the last layer of a convolutional neural network 700 can be a fully connected layer 715. The fully connected layer 715 is the connection layer between the preceding node layer 714 and the following node layer 716. The fully connected layer 713 is characterized by the fact that most, especially all, edges exist between nodes 714 of the preceding node layer 714 and nodes 716 of the following node layer, and that the weight of each of these edges can be adjusted individually.

[0152] In this embodiment, the nodes 724 of the preceding node layer 714 of the fully connected layer 715 are shown as a two-dimensional matrix, and additionally, as unrelated nodes, represented as a row of nodes, where the number of nodes is reduced for better presentation. This operation is also known as flattening. In this embodiment, the number of nodes 726 of the following node layer 716 of the fully connected layer 715 is less than the number of nodes 724 of the preceding node layer 714. Alternatively, the number of nodes 726 can be equal to or greater than the number of nodes 724.

[0153] Furthermore, in this embodiment, the Softmax activation function is used in the fully connected layer 715. By applying the Softmax function, the sum of the values ​​of all nodes 726 in the output layer 716 is 1, and all values ​​of all nodes 726 in the output layer 716 are real numbers between 0 and 1. In particular, if the input data is classified using the convolutional neural network 700, the values ​​of the output layer 716 can be interpreted as the probability that the input data falls into one of the different categories.

[0154] In particular, a convolutional neural network 700 can be trained based on the backpropagation algorithm. To prevent overfitting, regularization methods can be used, such as missing nodes 720, ..., 724, random pooling, using artificial data, weight decay based on L1 or L2 norm, or maximum norm constraint.

[0155] exist Figure 11 In the example shown, the CNN has a U-Net architecture. In the example illustrated, the input data for the CNN is a two-dimensional medical image containing 512x512 pixels, with each pixel containing an intensity value. The CNN contains convolutional layers (indicated by solid horizontal arrows), pooling layers (indicated by solid downward arrows), and upsampling layers (indicated by solid upward arrows). The corresponding number of nodes is indicated within the box. Within the U-Net architecture, the input image is first downsampled, specifically by reducing the image size and increasing the number of channels. Then, it is upsampled, specifically by increasing the image size and decreasing the number of channels, to generate the transformed image.

[0156] All convolutional layers except the last ones L1, L2, L4, L5, L7, L8, L10, L11, L13, L14, L16, L17, L19, L20 use a 3x3 kernel with padding value 1, a ReLU activation function and a number of filters or convolutional kernels adapted to the number of channels of the respective node layer as shown in Figure 11

[0157] The pooling layers L3, L6, L9 are max-pooling layers that replace four adjacent nodes by one node whose value is the maximum of the four adjacent nodes. The up-sampling layers L12, L15, L18 are transposed convolutional layers with a 3x3 kernel and a stride of 2, which actually increase the number of nodes by a factor of four. The dashed horizontal arrows correspond to a concatenation operation, where the output of the convolutional layers L2, L5, L8 of the down-sampling branch of the U-Net structure are used as additional input for the convolutional layers L13, L16, L19 of the up-sampling branch of the U-Net structure. The additional input data are considered as additional channels in the input node layer of the convolutional layers L13, L16, L19 of the up-sampling branch.

[0158] For training the CNN, a database containing 500 first medical images is used, wherein a respective segmentation mask is created based on annotations by an expert radiologist. In particular, the expert determines for each of the 500 first medical images a segmentation mask of the structure of interest, wherein a value of 1 is assigned to pixels corresponding to the structure of interest and a value of 0 is assigned to pixels not corresponding to the structure of interest. The database is divided into training data (320 data sets), validation data (80 data sets) and test data (100 data sets). For training the CNN, a backpropagation algorithm is used based on a binary cross-entropy cost function

[0159]

[0160] wherein

[0161] BCE(a,b) := - alog(b)(b) - (1 - a)log(l - b),

[0162] where x denotes a first medical image, y determines a corresponding segmentation mask created by an expert radiologist, and M(x) denotes the result of applying the CNN to the first input medical image x. Alternatively, other cost functions can be used, like weighted binary cross-entropy, Focal Loss or Dice Loss.

[0163] ​The best performing machine learning model is selected from several machine learning models (with different hyperparameters, such as number of layers, size and number of kernels, padding values, etc.) based on a validation set of 80 datasets and corresponding annotations. The specificity and sensitivity are determined based on a test set comprising 100 datasets and corresponding annotations.

[0164] In this patent application, pronouns and nouns referring to a person are generally not limited to a specific gender.

Claims

1. A computer-implemented method for characterizing perfusion defects in vascular structures of at least one organ (5, 8) of a patient based on imaging, wherein, - Receive medical imaging data (11, 12, 17) depicting the at least one organ (5, 8) and the vascular structure, wherein the medical imaging data (11, 12, 17) includes energy-resolved CT imaging data (11, 12, 17). - Detect blood flow obstruction objects (7) in the vascular structure based on the medical imaging data (11, 12, 17); and - Determine the perfusion defect fraction (16) of the target region (6a, 6b) of the at least one organ (5, 8) based on the energy-resolved CT imaging data (11, 12, 17), the blood perfusion of the target region potentially affected by the obstruction object (7).

2. The computer-implemented method according to claim 1, wherein, - Based on the medical imaging data (11, 12, 17), a segmentation (13) is generated to divide the at least one organ (5, 8) into multiple segments (6, 9), wherein the multiple segments (6, 9) are classified according to the blood supply to these segments by the vascular structures; and - The target region (6a, 6b) is identified as containing one or more target segments (6a, 6b) among the plurality of segments (6, 9), the blood perfusion of which is potentially affected by the obstruction object (7) according to the grading classification.

3. The computer-implemented method according to claim 2, wherein, For each of the one or more target segments (6a, 6b), a corresponding segment perfusion defect score is determined based on the energy-resolved CT imaging data (11, 12, 17), and the perfusion defect score (16) is determined based on the segment perfusion defect score.

4. The computer-implemented method according to claim 3, wherein, For each of the one or more target segments (6a, 6b), the size of the perfusion defect region (15) in the corresponding target segment (6a, 6b) is determined based on the energy-resolved CT imaging data (11, 12, 17), and the corresponding segment perfusion defect score is determined according to the size of the perfusion defect region (15).

5. The computer-implemented method according to any one of claims 2 to 4, wherein, The segmentation (13) is generated by applying a trained first machine learning model (MLM) to first input data that at least partially contains the medical imaging data (11, 12, 17).

6. The computer-implemented method according to any one of the preceding claims, wherein, At least one perfusion volume (PBV) value for the target region (6a, 6b) is determined based on the energy-resolved CT imaging data (11, 12, 17), and the perfusion defect fraction is determined based on the at least one PBV value.

7. The computer-implemented method according to claim 6 and any one of claims 2 to 5, wherein, The at least one PBV value includes the corresponding segment PBV value for the one or more target segments (6a, 6b).

8. The computer-implemented method according to any one of the preceding claims, wherein, The medical imaging data (11, 12, 17) includes photon-counted CT (PCCT) imaging data and the obstruction object (7) is detected based on the PCCT imaging data (11, 12, 17).

9. The computer-implemented method according to any one of the preceding claims, wherein, The obstructed object (7) is detected by applying a trained second MLM to second input data that at least partially contains the medical imaging data (11, 12, 17).

10. The computer-implemented method according to any one of the preceding claims, wherein, The energy-resolved CT imaging data (11, 12, 17) include contrast-enhanced CT (CECT) imaging data (12).

11. The computer-implemented method according to any one of the preceding claims, wherein, The perfusion defect score (16) is determined by applying the trained third MLM (10) to the third input data (11, 12, 13) containing the medical imaging data (11, 12, 17).

12. The computer-implemented method according to any one of the preceding claims, wherein, The at least one organ (5, 8) includes the patient's lungs.

13. A data processing system (4) configured to perform a computer-implemented method according to any one of the preceding claims.

14. A medical imaging system comprising a data processing system (4) according to claim 13 and a CT device (2), the CT device being configured to generate medical imaging data (11, 12, 17) depicting the at least one organ (5, 8) and the vascular structures.

15. A computer program product comprising instructions that, when executed by a data processing system (4), cause the data processing system (4) to perform a computer-implemented method according to any one of claims 1 to 12.

Images