Machine learning based image segmentation

Abstract

A system (TFS) and related method for processing training data for facilitating training of a machine learning model (M) in a segmentation task. The system may comprise an input interface (IN) through which is receivable training data (e), such training data comprising spatial data and capable of representing i) instances of inimage objects and ii) background. A mask generator (MG) generates masks (m) for the training data, wherein a given such mask pertains to a single such instance of the said objects, referred to herein as an object instance, and further pertains to only a subset of the background surrounding the said object instance. An output interface (OUT) provides the so generated masks to facilitate the of training of the machine learning model.

Description

[0001]Brainlab AG Attorney’s File: B19045WO / AO MACHINE LEARNING BASED IMAGE SEGMENTATION FIELD OF THE INVENTIONThe present invention relates to system for processing training data for facilitatingtraining of a machine learning model in a segmentation task, to a training system fortraining a machine learning model, to a machine learning based system, to relatedmethods, to uses of the processed training data, to a corresponding computerprogram, to a non-transitory program storage medium storing such a program, and toa computing system for executing the program.TECHNICAL BACKGROUND Medical imaging is one of the most useful tools in the medic’s arsenal. Being able to see inside a patient in a non-invasive manner informs in a meaningful way therapy and / or diagnoses. Imaging modalities have been growing ever more sophisticated, from a humble beginning in the 19thcentury as x-ray projection imaging (still much used to this day), to tomographic imaging, such as X-ray based CT imaging in the 50’s, nuclear imaging such as PET / SPECT in the 60’s, and later the addition of magnetic resonance imaging (MRI) in the 70’s. The blessings of those powerful imaging technologies pose perhaps unexpected challenges: imagery becomes more and more complex to work with. Radiologists are required to review ever more complex sets of data such as tomographic image volumesthat are made up of hundreds of slices, some of them having thicknesses in the sub-millimeter range. Add to this the ability to reformat views on these slices along differentspatial directions, yields, to some, a bewildering landscape to navigate. It can make a targeted efficient review by a human reviewer challenging, and may slow down output.In a 2022 paper published in Radiology, vol 304, pp 274-282 by R. Alexander et al as“Mandating Limits on Workload, Duty and Speed in Radiology” it was estimated thatthe average CT volume includes about 679 individual cross-sectional images (“slices”).On an average working day, with the current workload, this may leave a little more than a second for the radiologist to look at each image slice they are reporting on. In addition, in some departments, it is the workload itself that has increased dramatically by more than 500% as reported in the cited paper. And it is also the aging population in some parts of the world that puts strain on radiology departments, with a workload difficult to manage. The combination of these factors, with the availability of more complex imagery has been found to result in an increase in the number of reading errors. Consequences of such reading errors can result from the inconvenient and expensive where unwarranted treatments are administered, to the sheer catastrophic where patients succumb before their time to undetected disease. Similar effects may be felt across other image-based medical endeavours, such as in radiotherapy or surgery planning, and others. This need has been recognised by providing across the industry computerized systems. Such systems may assist, for example radiologists, in image reading tasks. At its core, some such computerized systems ease the burden on the clinical user by being capable of automated analysis of imagery. They extract information or distil theinformation residing therein. This may help human reviewer find relevant informationquicker, such as in diagnostic tasks (eg, screening programs), image-based navigation, or planning, such as for radiotherapy, or other. One such computer-aided information extraction is image segmentation. Image segmentation may be conceptualized as an exercise in classifying image elements, such as pixels or voxels of which the imagery is made up, into different classes of interest. For example, a binary segmentation may be classifying, for each pixel, whether it represents part of a cancerous lesion (one class) or not (another class). Thus, the imagery is essentially broken up into discrete parts (“segments”), of which some may represent cancerous tissues, whilst others are estimated by the segmentation to represent non-cancerous tissue. More advanced, multi-class classification can also be done. The image so broken up can then be visualized for example where different segments (and hence classes) appear in different colours, shadings or in other visual modulations. Thus, segmentation may help user to quicklypin-point any suspicious masses or structures. Segmentation may help radiologist tointerpret (“read”) imagery, with greater certainty, to rule out or establish disease. One use case may be breast cancer screening, where x-ray imagery is analysed for image structures betoken of nodules, lesions that may develop into cancer. Previously, image segmentation was done using analytic-classical methods, where computation was based on sometimes closed-form expression, etc, to perform thesegmentation: snakes, region growing or thresholding are examples. However, theseanalytical methods, useful as they may have been and still are for some cases, suffered from shortcomings, when it came to more complex distinctions to be drawn. Robustness was often of want. A more recent segmentation approach uses machinelearning (“ML”). ML models are used in such approaches. They are trained on a largenumber of prior training data such as imagery. This training allows building more flexible, more robust, segmentation tools. However, in training, certain challenges have been observed, such as the data imbalance problem. This imbalance may come about in the training data when there is not enough data examples to go around for each ofthe intended classes into which one may wish to segment the imagery for: there maybe much more examples of training data for one class than there are for another. This can lead to under-performing machine learning models once trained on such imbalanced training data set. Another imbalance problem is intra-class imbalance where objects within the same class may differ considerably in scale. Thus, there may therefore be a need for improved approaches in training machine models for processing spatial data, such as segmenting such data, in particular image data, yet more particularly, medical image data.Aspects of the present invention, examples and exemplary steps and theirembodiments are disclosed in the following. Different exemplary features of the invention can be combined in accordance with the invention wherever technically expedient and feasible. EXEMPLARY SHORT DESCRIPTION OF THE INVENTION In the following, a short description of the specific features of the present invention is given which shall not be understood to limit the invention only to the features or a combination of the features described in this section. What is proposed herein is a system and method to facilitate training a machine learning model, in particular for a segmentation task to segment for in-image objects of certain classes.The proposed system and method allow addressing the (training) data imbalanceproblem that, if not addressed, might undermine performance of the trained model when trained on such imbalanced training data. The proposed method and system processes training data to produce masks for the training data that can be used in evaluating the objective function, such a loss function, over the training data set.The proposed masks allow addressing intra-class imbalance, where object instanceswithin the same class may come in different sizes. Some training may include imagerythat may show small and large lesions at the same time. If not accounted for, theobjective function may apportion too little a contribution for smaller object instances, giving undue contribution to larger object instances. For example, in a loss functionformulation, the loss contribution of the small lesion would be small, and the smalllesion would potentially be missed by the trained model.The masks generated herein allow focussed processing of the training data, with , inparticular, special handling of background information, which may otherwise lead tothe said inter-class imbalance. The masks are dedicated, focussed, on a single instance of an object at time, whilst the background is processible separately by way of another mask more focused on the said background. In addition, because of the dedicated single object instance masks, each such mask is considered separately by the objective function (eg, loss function) to obtain estimates for partial losses, one such partial loss per object instance loss and one such partial loss measurement for background. The sum over the partial losses then yields the total sum (total loss). The model parameters are adjusted in the training to improve (lower) the total loss. All such partial losses may be weighted equally, which disencumbers the proposed setup from design elaborate weighing schemes. The proposed mask allowed improved more useful consideration of background information with a reduction of contradicting signals considered by the model, and by the objective function, with background information contributing to the evaluation of the objective function only when necessary, rather than multiple times over. Objects of interest are considered when it is their turn, whilst other information is essentially masked out. The proposed method and system can be used with a range of objective functions such as cost functions, in particular of the Dice-type. GENERAL DESCRIPTION OF THE INVENTION The following will be confined chiefly to applications in the medical realm, but applications outside medical, such as in engineering are also envisaged herein, where spatial data of whichever type is required and where there is a need for analysis ofsuch data.In general, the aforementioned needs are met by providing, in a first aspect, a systemfor processing training data for facilitating training of a machine learning model in asegmentation task, comprising: a) an input interface through which is receivable training data (e), such training data comprising spatial data and capable of representing i) instances of in-image objects and ii) background; b) a mask generator capable of generating masks for the training data, wherein a given such mask pertains to a single such instance of the said objects, referred to herein as an object instance, and further pertains to only a subset of the background surrounding the said object instance, andc) an output interface for providing the so generated masks to facilitate the trainingof the machine learning model.In embodiments, the mask generator is capable to so generate the given mask basedon a respective prior mask for the said object instance, wherein such prior mask pertains only to the said object instance, at the exclusion of background, and wherein the generating of the mask by generator is based on expanding the prior mask to then include the said subset of background.In embodiments, the mask generator is capable to so generate a further mask, thatpertains to the, or another, object instance, and further pertains to the whole of the background or to a subset thereof that is larger than the surrounding background subset. In embodiments, the said object instance is representative of a clinical feature. In embodiments, the training data includes imagery. In embodiments, the said imagery is medical imagery. In embodiments, the said spatial data is providable by an imaging modality. In embodiments, the said imaging modality includes one of a medical imager, and a surveillance camera.In a second aspect there is provided a training system for training a machine learningmodel based on the masks provided by the training data provider system as per any one of the preceding claims.In embodiments, the training system is driven by an objective function wherein the saidobjective function is configured for per mask processing of the training data. The facilitator system as per the first aspect or any one of its embodiments, may bepart of the objective function or entity (call, routine, module, etc) that executes theobjective function. Some or each such mask so generated specifies admissionzones / areas in the spatial data (eg, image) that determine which part of the spatialdata is to be processed by objective function, and which is to be excluded from soprocessing. Such control of processing may be the case for binary (“hard”) masks.However, “soft”, or “fuzzy”, masks specifying weights are also envisaged where thepart of the spatial data outside the admission zone is down-weighted relative to spatialdata inside designation zone In embodiments, the said objective function is based on loss or utility per mask. In embodiments, the loss function is any one of Dice-based loss, or Focal-loss.In a third aspect, there is provide a machine learning based system including themachine learning model as trained by the system as provided as per the secondaspect, or any one of its embodiments. In embodiments, machine learning based system is any one of medical, or surveillance. In embodiments, the machine learning based system is capable of instance segmentation.In a fourth aspect, there is provided a computer-implemented method for processingtraining data for training a machine learning model in a segmentation task, comprisingthe following steps: i) receiving training data, such training data comprising spatial data and capable of representing i) instances of in-image objects and ii) background; ii) generating masks for the training data, wherein a given such mask pertains to a single such instance of the said objects, referred to herein as an object instance, and further pertains to only a subset of the background surrounding the said object instance, and iii) providing the so generated masks to facilitate the of training of the machine learning model.In a fifth aspect there is provided a method for training a machine learning model in asegmentation task, based on training data processable by the method as provided asper the fourth aspect.In a sixth aspect there is provided a method of inference or testing, based on a machinelearning model trained on training data processable by the method as provided as perthe fourth aspect.In a seventh aspect there is provided a use of masks for training data in training amachine learning model, the said masks obtainable by processing by the method as provided as per the fourth aspect.In an eight aspect there is provided a use of a machine learning model trained basedon training data having masks applied as obtainable by processing the training data by the method as provided as per the fourth aspect. In a yet further aspect, the invention is directed to a computer program which, when running on at least one processor (for example, a processor) of at least one computer (for example, a computer) or when loaded into at least one memory (for example, amemory) of at least one computer (for example, a computer), causes the at least onecomputer to perform any one of the above-described methods. The invention may alternatively or additionally relate to a (physical, for example electrical, for example technically generated) signal wave, for example a digital signal wave, carrying information which represents the program, for example the aforementioned program, which for example comprises code means which are adapted to perform any or all ofthe steps of any one of the above-mentioned methods. A computer program stored ona disc is a data file, and when the file is read out and transmitted it becomes a data stream for example in the form of a (physical, for example electrical, for exampletechnically generated) signal. The signal can be implemented as the signal wave whichis described herein. For example, the signal, for example the signal wave is constitutedto be transmitted via a computer network, for example LAN, WLAN, WAN, for examplethe internet. The invention according to the second aspect therefore may alternativelyor additionally relate to a data stream representative of the aforementioned program. In a third aspect, the invention is directed to a non-transitory computer-readable program storage medium on which the program according to the fourth aspect is stored.In a yet further aspect, the invention is directed to at least one computing system (forexample, a computer), comprising at least one data processor (for example, a microprocessor) and at least one memory (for example, a volatile or non-volatile memory), wherein the program according to the fourth aspect is running on the processor or is loaded into the memory, or wherein the at least one computer comprises the computer-readable program storage medium according to the fifth aspect. The proposed systems and methods address a need in machine learning. Specifically, in machine learning, machine learning models are trained using an objective function, also called a loss function in some contexts, that measures how well the model performs by comparing the model's prediction to the ground-truth, i.e., true labels. One of these losses used herein in embodiments is the Dice loss, which is modified herein in order to allow the model to better detect and segment in-image objects of interest (such as representations of lesions, etc) of different sizes in patient scans for example.The loss function used herein, may be called "Masked Instance Dice Loss", may beused to train a lesion segmentation model or a model for other tasks. It showedimproved performance in comparison to other state-of-the-art losses. The masked instance Dice loss computes the Dice loss separately for some or every object instance, where instances in lesion detection may relate to expert-annotated lesions, or in whichever manner the ground truth information was gotten. The proposed setup creates a mask per object / lesion, may then expand this mask to include some of the background region surrounding the lesion object, thereby allowing for a better delineation of the lesion object. The “partial” loss for a particular instance is then defined as the negative Dice score of voxels / pixels only within the region defined by the mask. In addition to computing the dice loss for every lesion object, the proposed setup also takes into account contribution of the background class by constructing a dedicated background (“combo") mask that combines one of the lesion objects, with the background segmentation (map), and computes the Dice for voxels inside thismask. The final loss may then be computed as the average loss of N+1 components,where N is the number of lesion objects in the ground-truth segmentation, and the extracomponent is combo mask. Given that each instance contributes equally to the loss, this loss is particularly helpful for applications where objects (such as lesions) can vary significantly in their sizes. Since the loss gives equal weighting to each instance, small lesions can be segmented as well as larger ones. Furthermore, as the proposed loss has a notion of instances, the resulting models are better equipped to not only segment lesions correctly, but also to detect the correct number of lesions in an image. It was found that the proposed approach is generic, in particular is objective function- invariant, and can hence be applied to any objective function, such as any other loss / cost or utility function. The proposed setup allows for improved consideration of background in the imagery. There are no contradicting signals. Background voxels / pixels contribute tothe loss when necessary, rather than multiple times over as in previous approaches.The image objects (eg, representative of lesions in medical imagery) are only considered when it is “their turn” for a given partial loss in the overall summation in the loss function. Other objects are masked out for any given partial loss (summand) in the sum of losses that form the total loss as measured by the loss function. The described manner of constructing the masks, each mask pertains to a single instance of object. Thus, each mask may be said to represent a focus window, focused on a single such instance of the object of interest. In particular, it is the processing by the loss function F of the training imagery that is so focused. Each mask may define an admission area and an exclusion are. The admission area defined by each such mask is such that the background therein is smaller, indeed much smaller, than the area occupied by the object instance. Thus, the background area in each of the object instance focused mask’s admission area is small in comparison to other approaches where the masks there are such that it is the background that occupies the larger area. With the proposed mask, the admitted portion of the background is also much smaller than the entire background that is initially present. Thus, with the proposed focused masks, with small proportion of admission area taken up by background, a contribution or “weight” of the background in the processing byobjective function (eg loss function) is reduced. Such masks allow an improvedweighing scheme of loss contributions of each instance in respect to the overall lossmeasured by loss function. In particular, the proposed mask generation scheme allows weighing each loss contribution equally. Thus, there is no need to construct involved weighing schemes as hyper-parameters. In the past, having to construct such weight distributions as hyper-parameters consumed much time and costs. No efforts need tobe spent on this in the proposed approach, where each instance / mask attracts thesame weight. This is straightforward to implement.It will be understood that the masks may be generated by proceeding from a given pre-segmented instance of a lesion and expanding a margin around this. The margin width may be constant for each masking instance preferably, although the area sizes, the admitted area AA may differ depending on the size of the lesion / pre-segmented object OJ. Thus, whilst the margin itself may be constant across masks the background itself may differ and is larger in general for larger lesions. It will be understood that the use of “size” as used herein may depend on the dimensionality of the underlying imagery and hence mask. Thus, the masks and the imagery may be two three or four dimensional as needed and therefore the size here may not necessarily be a 2D area size but may in fact be a volume. Consequently, higher than two dimensional masks as shown in Figure 6 may comprise a cube, ellipsoid or any other neighborhood structure as needed defining a volume with exclusion areas, volumes defined therein. The mask as generated herein may be applied to the training data, in particular ground truth and the predictions output by model during training, in particular during evaluation by objective function. The masks may be applied to the training data upfront to obtainmasked training data items, and these are then passed for processing to trainingsystem, in particular to objective function. Alternatively, the masks are applied as data“filters” in the evaluation by objective function or by a related logic, on-the-fly. Thus,the masked may be considered to mask the objective function, thereby providing a masked objective function, such as a “masked-Dice-loss (function)”, etc. Also, it will be understood that in the present disclosure the term object instance may be read in connection with the complementary term of background, the background comprising everything that is not an instance of an object.Such object ^ considered as a sub-set of pixels may be a contiguous, topologicallyconnected, in-image structure, and such structure is preferred herein, this may notnecessarily be so in all embodiments. Such in-image structure may represent a pathological physical structure, such as a lesion, or other. The invention does not involve or in particular comprise or encompass an invasive step which would represent a substantial physical interference with the body requiring professional medical expertise to be carried out and entailing a substantial health risk even when carried out with the required professional care and expertise. For example, the invention does not comprise a step of positioning a medical implant in order to fasten it to an anatomical structure or a step of fastening the medical implant to the anatomical structure or a step of preparing the anatomical structure for having the medical implant fastened to it. More particularly, the invention does not involve or in particular comprise or encompass any surgical or therapeutic activity. The invention isinstead directed as applicable to facilitating the training of a machine learning model.For this reason alone, no surgical or therapeutic activity and in particular no surgical or therapeutic step is necessitated or implied by carrying out the invention. Use of systems and methodsThe present invention also relates to the use of the systems and methods, and anyembodiments thereof, for segmentation of spatial data, in particular of image data, andmore particularly still of medical image data. DEFINITIONSThis section lists definitions for terminology as is used in this disclosure, and whichform part of the present disclosure. Computer-implemented Method The method in accordance with the invention is for example a computer implementedmethod. For example, all the steps or merely some of the steps (i.e. less than the totalnumber of steps) of the method in accordance with the invention can be executed by a computer (for example, at least one computer). An embodiment of the computer implemented method is a use of the computer for performing a data processing method. An embodiment of the computer implemented method is a method concerning the operation of the computer such that the computer is operated to perform one, more or all steps of the method. The computer for example comprises at least one processor and for example at least one memory in order to (technically) process the data, for example electronically and / or optically. The processor being for example made of a substance or composition whichis a semiconductor, for example at least partly n- and / or p-doped semiconductor, forexample at least one of II-, III-, IV-, V-, VI-semiconductor material, for example (doped) silicon and / or gallium arsenide. The calculating or determining steps described are for example performed by a computer. Determining steps or calculating steps are for example steps of determining data within the framework of the technical method, for example within the framework of a program. A computer is for example any kind of data processing device, for example electronic data processing device. A computer can be a device which is generally thought of as such, for example desktop PCs, notebooks, netbooks, etc., but can also be any programmable apparatus, such as for example a mobile phone or an embedded processor. A computer can for example comprise a system (network) of "sub-computers", wherein each sub-computer represents a computer in its own right. The term "computer" includes a cloud computer, for example a cloud server. The term "cloud computer" includes a cloud computer system which for example comprises a system of at least one cloud computer and for example a plurality of operatively interconnected cloud computers such as a server farm. Such a cloud computer is preferably connected to a wide area network such as the world wide web (WWW) and located in a so-called cloud of computers which are all connected to the world wide web. Such an infrastructure is used for "cloud computing", which describes computation, software, data access and storage services which do not require the end user to know the physical location and / or configuration of the computer delivering a specific service. For example, the term "cloud" is used in thisrespect as a metaphor for the Internet (world wide web). For example, the cloudprovides computing infrastructure as a service (IaaS). The cloud computer can function as a virtual host for an operating system and / or data processing application which is used to execute the method of the invention. The cloud computer is for example an elastic compute cloud (EC2) as provided by Amazon Web Services™. A computer for example comprises interfaces in order to receive or output data and / or perform an analogue-to-digital conversion. The data are for example data which represent physical properties and / or which are generated from technical signals. The technical signals are for example generated by means of (technical) detection devices (such as for example devices for detecting marker devices) and / or (technical) analytical devices (such as for example devices for performing (medical) imaging methods), wherein the technical signals are for example electrical or optical signals. The technical signals for example represent the data received or outputted by the computer. The computer is preferably operatively coupled to a display device which allows information outputted by the computer to be displayed, for example to a user. One example of a display device is a virtual reality device or an augmented reality device (also referred to as virtual reality glasses or augmented reality glasses) which can be used as "goggles" for navigating. A specific example of such augmented reality glasses is Google Glass (a trademark of Google, Inc.). An augmented reality device or a virtual reality device can be used both to input information into the computer by user interaction and to display information outputted by the computer. Another example of a display device would be a standard computer monitor comprising for example a liquid crystal display operatively coupled to the computer for receiving display control data from the computer for generating signals used to display image information content on the display device. A specific embodiment of such a computer monitor is a digital lightbox. An example of such a digital lightbox is Buzz®, a product of Brainlab AG. The monitor may also be the monitor of a portable, for example handheld, device such as a smart phone or personal digital assistant or digital media player. The invention also relates to a program which, when running on a computer, causes the computer to perform one or more or all of the method steps described herein and / or to a program storage medium on which the program is stored (in particular in a non- transitory form) and / or to a computer comprising said program storage medium and / or to a (physical, for example electrical, for example technically generated) signal wave, for example a digital signal wave, carrying information which represents the program, for example the aforementioned program, which for example comprises code means which are adapted to perform any or all of the method steps described herein. Within the framework of the invention, computer program elements can be embodied by hardware and / or software (this includes firmware, resident software, micro-code, etc.). Within the framework of the invention, computer program elements can take the form of a computer program product which can be embodied by a computer-usable, for example computer-readable data storage medium comprising computer-usable, for example computer-readable program instructions, "code" or a "computer program" embodied in said data storage medium for use on or in connection with the instruction- executing system. Such a system can be a computer; a computer can be a data processing device comprising means for executing the computer program elements and / or the program in accordance with the invention, for example a data processing device comprising a digital processor (central processing unit or CPU) which executes the computer program elements, and optionally a volatile memory (for example a random access memory or RAM) for storing data used for and / or produced by executing the computer program elements. Within the framework of the present invention, a computer-usable, for example computer-readable data storage medium can be any data storage medium which can include, store, communicate, propagate or transport the program for use on or in connection with the instruction-executing system, apparatus or device. The computer-usable, for example computer-readable data storage medium can for example be, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared or semiconductor system, apparatus or device or a medium of propagation such as for example the Internet. The computer-usable or computer-readable data storage medium could even for example be paper or another suitable medium onto which the program is printed, since the program could be electronically captured, for example by optically scanning the paper or other suitable medium, and then compiled, interpreted or otherwise processed in a suitable manner. The data storage medium is preferably a non-volatile data storage medium. The computer program product and any software and / or hardware described here form the various means for performing the functions of the invention in the example embodiments. The computer and / or data processing device can for example include a guidance information device which includes means for outputting guidance information. The guidance information can be outputted, for example to a user, visually by a visual indicating means (for example, a monitor and / or a lamp) and / or acoustically by an acoustic indicating means (for example, a loudspeaker and / or a digital speech output device) and / or tactilely by a tactile indicating means (for example, a vibrating element or a vibration element incorporated into an instrument). For the purpose of this document, a computer is a technical computer which for example comprises technical, for example tangible components, for example mechanical and / or electronic components. Any device mentioned as such in this document is a technical and for example tangible device. Acquiring Image Data The expression "acquiring data" for example encompasses (within the framework of a computer implemented method) the scenario in which the data are determined by the computer implemented method or program. Determining data for example encompasses measuring physical quantities and transforming the measured values into data, for example digital data, and / or computing (and e.g. outputting) the data by means of a computer and for example within the framework of the method in accordance with the invention. The meaning of "acquiring data" also for example encompasses the scenario in which the data are received or retrieved by (e.g. inputto) the computer implemented method or program, for example from another program,a previous method step or a data storage medium, for example for further processing by the computer implemented method or program. Generation of the data to be acquired may but need not be part of the method in accordance with the invention. The expression "acquiring data" can therefore also for example mean waiting to receive data and / or receiving the data. The received data can for example be inputted via an interface. The expression "acquiring data" can also mean that the computer implemented method or program performs steps in order to (actively) receive or retrieve the data from a data source, for instance a data storage medium (such as for example a ROM, RAM, database, hard drive, etc.), or via the interface (for instance, from another computer or a network). The data acquired by the disclosed method or device, respectively, may be acquired from a database located in a data storage device which is operably to a computer for data transfer between the database and the computer, for example from the database to the computer. The computer acquires the data for use as an input for steps of determining data. The determined data can be output again to the same or another database to be stored for later use. The database or database used for implementing the disclosed method can be located on network data storage device or a network server (for example, a cloud data storage device or a cloud server) or a local data storage device (such as a mass storage device operably connected to at least one computer executing the disclosed method). The data can be made "ready for use" by performing an additional step before the acquiring step. In accordance with this additional step, the data are generated in order to be acquired. The data are for example detected or captured (for example by an analytical device). Alternatively or additionally, the data are inputted in accordance with the additional step, for instance via interfaces. The data generated can for example be inputted (for instance into the computer). In accordance with the additional step (which precedes the acquiring step), the data can also be provided by performing the additional step of storing the data in a data storage medium (such as for example a ROM, RAM, CD and / or hard drive), such that they are ready for use within the framework of the method or program in accordance with the invention. The step of "acquiring data" can therefore also involve commanding a device to obtain and / or provide the data to be acquired. In particular, the acquiring step does not involve an invasive step which would represent a substantial physical interference with the body, requiring professional medical expertise to be carried out and entailing a substantial health risk even when carried out with the required professional care and expertise. In particular, the step of acquiring data, for example determining data, does not involve a surgical step and in particular does not involve a step of treating a human or animal body using surgery or therapy. In order to distinguish the different data used by the present method, the data are denoted (i.e. referred to) as "XY data" and the like and are defined in terms of the information which they describe, which is then preferably referred to as "XY information" and the like. Imaging In the field of medicine, imaging methods (also called imaging modalities and / or medical imaging modalities) are used to generate image data (for example, two- dimensional or three-dimensional image data) of anatomical structures (such as soft tissues, bones, organs, etc.) of the human body. The term "medical imaging methods" is understood to mean (advantageously apparatus-based) imaging methods (for example so-called medical imaging modalities and / or radiological imaging methods) such as for instance computed tomography (CT) and cone beam computed tomography (CBCT, such as volumetric CBCT), x-ray tomography, magnetic resonance tomography (MRT or MRI), conventional x-ray, sonography and / or ultrasound examinations, and positron emission tomography. For example, the medical imaging methods are performed by the analytical devices. Examples for medical imaging modalities applied by medical imaging methods are: X-ray radiography, magnetic resonance imaging, medical ultrasonography orultrasound, endoscopy,elastography, tactile imaging, thermography, medicalphotography and nuclear medicine functional imaging techniques as positronemission tomography (PET) and Single-photon emission computedtomography (SPECT), as mentioned by Wikipedia.The image data thus generated is also termed “medical imaging data”. Analytical devices for example are used to generate the image data in apparatus-based imaging methods. The imaging methods are for example used for medical diagnostics, to analyze the anatomical body in order to generate images which are described by the image data. The imaging methods are also for example used to detect pathologicalchanges in the human body. However, some of the changes in the anatomicalstructure, such as the pathological changes in the structures (tissue), may not be detectable and for example may not be visible in the images generated by the imagingmethods. A tumor represents an example of a change in an anatomical structure. If thetumor grows, it may then be said to represent an expanded anatomical structure. Thisexpanded anatomical structure may not be detectable; for example, only a part of the expanded anatomical structure may be detectable. Primary / high-grade brain tumors are for example usually visible on MRI scans when contrast agents are used to infiltrate the tumor. MRI scans represent an example of an imaging method. In the case of MRIscans of such brain tumors, the signal enhancement in the MRI images (due to thecontrast agents infiltrating the tumor) is considered to represent the solid tumor mass.Thus, the tumour is detectable and for example discernible in the image generated by the imaging method. In addition to these tumors, referred to as "enhancing" tumors, itis thought that approximately 10% of brain tumors are not discernible on a scan andare for example not visible to a user looking at the images generated by the imaging method. Machine Learning (“ML”) includes implementing a machine learning algorithm, on any computing entity. Some such ML algorithms operate to adjust parameters of a machine learning model, thusconfiguring same to perform (“learn”) a task. This adjusting or updating of the model iscalled “training”. In general, performance of the ML model in respect of the task may improve measurably with training experience. Training experience may include suitable training data, and exposure of the model to such data. Task performance may improve, the better the data represents the task to be learned, and the more different samples of such training data are available. The performance may be measured by objective tests based on output produced by the model in response to feeding the model with test data. The performance may be defined in terms of a certain error rate to be achieved. See for example, T. M. Mitchell, “Machine Learning”, McGraw-Hill, 1997. Segmentation is an operation on spatial data (such as imagery), where elements (such as pixels or voxels) are classified into two or more classes. Segmentation includes semantic segmentation. Segmentation may be implemented using ML. Segmentation may also be formulated as a regression problem, rather than aclassification problem. However, formulation of segmentation as a classificationproblem is preferred herein. Pixels Any reference herein to “pixels” is merely synecdochic shorthand and may refer also to voxels, or other image elements such as sets (chunks) of pixels or voxels, depending on the level of granularity intended in the processing. The proposed methods andsystem are capable of operating on per pixel / voxel level, or on such chunks.Size / Area and related terms, as used herein, may depend on the dimensionality of the underlying data. Thus, any reference herein to “area”, such as “admission area”, etc, ought to be read synecdochally, and may stand not only for 2D area, but for 3D volume or higher dimensional volume such as 4D or higher still, as needed. Similar, reference to “mask” and “masked image”, etc, may be sub-volumes in three or higher dimension, rather than 2D areas, depending on the dimension of the imagery. Mask as used herein is a function that is capable of controlling the manner in which spatial data, such as an image, is to be processed, in particular by an objective function. Suchobjective function as may be used in an ML training algorithm to drive the algorithmformulated as an optimization. The mask can have a (intrinsic) size. The size may be “spatial size” (in terms of pixels of voxels). In addition or instead, the said size pertains to “focus window” for processing by the objective function or other entity. Thus, some areas on the spatial data may be ignored by the objective function or down-weighted, whilst other data is processed as they are or are up-weighted and then processed. Object as used herein is an in-image object, such as a structure that may relate to a physical structure, such as a lesion. The segmentation may be directed to such objects and may be configured to segment objects into one of plural classes. There may be multipleinstances of objects of the same class, and the segmentation may be configured tosegment different instances of objects from the same class (instance-based segmentation). “Object instances” or similar are instances of objects, distinct objects that a segmented into the same class. A given object may be topologically connected, preferably with no holes, although in some cases such holes may be present. Thus,an object instance may or may not be “simply connected” in the topological sense. Acollection of segmented objects, as opposed to background, may be understood asinstances of foreground.Optimization as used herein may or may not include attaining a global extremum (minimum or optimum). Attaining instead a local extremum may be sufficient. Atting such local orglobal extremum may include so attaining with a error margin. In any case, anoptimization is considered herein as numerically accomplished if a stopping condition is met. BRIEF DESCRIPTION OF THE DRAWINGSIn the following, the invention is described with reference to the appended figures whichgive background explanations and represent specific embodiments of the invention.The scope of the invention is however not limited to the specific features disclosed inthe context of the figures, whereinFig. 1 shows a block diagram of a data processing arrangement;Fig. 2 illustrates principles of data acquisition;Fig. 3 illustrates various manners of image segmentation;Fig. 4 shows a block diagram of a system for facilitating training of amachine learning model by processing training data;Fig.5 is an illustration of the concept of segmentation mask as may beused herein in embodiments;Fig. 6 illustrates operation of the system of Figure 4; andFig. 7 shows a flow-chart of a computer-implemented method offacilitating training of a machine learning based on training data. DESCRIPTION OF EMBODIMENTS Reference is first made to the block diagram of Figure 1, which shows a computer assisted data processing arrangement DPA. The arrangement DPA may be used in a range of medical applications medical, but use outside the medical field is not excludedherein, such as in image-based surveillance or other. However, for the sake ofdefiniteness, main reference is made herein to medical applications. The arrangement is mainly envisaged for handling spatial data e.Such spatial data e may be provided by a measurement device, such as a medicalimaging modality IA. Thus, in embodiments mainly envisaged herein, the spatial data is the form of medical imagery e, either still or video. The medical imagery is made up of image values e={v}, such as voxel values or pixel values. As an example and shorthand, such image values will be referred to herein simply as “pixels”, which is not at the exclusion of such values being volumetric voxel values or other measurements.Envisaged modalities IA include any one of X-ray based imaging such as Radiographicprojection imaging, computed tomography (CT), MRI, nuclear imaging such as PET, SPECT. Other tomographic modalities such as OCT (optical coherence tomography) are also included, and so is ultrasound (US), or any other.The imagery e provided by imager IA in an image acquisition may be forwardedthrough a wireless and / or wired communication system, such as a communication network (LAN, WLAN, WWW, Intranet, etc) or via any other, to one or more data consumers DC. Such data consumer DC may be a computer system CS, such a workstation, mobile or desktop, that may be used by a radiologist in an image reading session. The computer system CS may include a visualizer VIZ that allows visualizing the imageryon a display device DD. The visualizer VIZ maps pixel values to a suitable colour orgrey value palette and interacts with video circuitry of the display device to so cause agraphics display of the imagery e to be displayed on a screen of the display device DD.The radiologist may visually review the displayed imagery to compile a report with their findings which may inform therapy or diagnosis. Alternatively, the displayed imagery is reviewed by a planner to identify a target region for planning of a radiotherapy or a surgical procedure, for any other image-based task at hand An image processer IP of system CS may be used for computer-assisted support of user, or for automated analysis of the imagery, with results being displayed on displaydevice DD, stored or otherwise used. The image processer may process raw imagerye, and it is the processed imagery that is then visualized on the display device, asneeded. The image data e may be stored in volatile on non-volatile memory MEM. Anexample is a PACS or other medical data storage on which the data may be stored and retrieved when needed. The computing system CS may be part of a HIS (hospital information system). In embodiments, image processor IP is configured to support a radiologist in readingthe imagery so as to establish case for a disease. The image processor IP may beconfigured as an automated diagnosis decision system that processes the imagery and outputs a prediction on the presence of a dieses, condition, etc. In other cases, the image processor may automatically identify potentially suspicious image structures that may represent lesions. The computing CS may run on a single or plural computing entities or nodes. Distributed computing may be used, in an edge computing or Fog / Cloud computing setup. A client-server approach may be used. The image processor IP, or rather its capabilities, may be contracted in as a service. The lower part of Figure 1 reveals more details of the image processer / image processing entity (eg, such as a service) IP envisaged herein. The image processer IP (for such it will be referred to herein, regardless of the computing infrastructure used), may be based on, or may include, a machine learning system MS. Such ML systemMS uses machine learning. Specifically, it is operable, based on a (one or more) priortrained machine learning model M. Such model M is trained on training data. The machine learning model M may be trained to implement a segmentation operation σ. Thus, the image processor IP may be operable, thanks to its trained ML model M, tofunction as an image segmentation tool or service. The segmentation operation σ maybe a semantic segmentation operation. As mentioned earlier, segmentation operationσ may be conceptualized as a global or local image classification operation. Thesegmentation operation σ break downs, possibly complex, high dimensional,volumetric, imagery e into components (“segments´”) of interest. Such componentsmay be referred to as in-image objects They are representative of aspects, features, properties, etc of certain physical objects of interest, such physical objects being the purpose for the imaging examination. Before it can be used for segmentation in real- clinical practice (a phase sometimes called deployment or inference), the machine learning model needs to be prepared for this use in prior one or more phases, often referred to as “training” or “learning” phase, or, as will be done herein, simply as “training (procedure)”. In such training, the model M is trained on training data by a computing system, referred to as training system TS. This system TS is in general different from the “clinical use” system CS, as used in deployment, although the same computing system may be used in some instances. In general, the computation performance of thetraining system TS is higher than that of deployment system CS. This is because training can be a computationally demanding process. Quick access memory may be used in system TS, in addition or instead of specialist microprocessors, such as GPUsor multi-core processors, capable of parallel computing.Training procedure is based on training data Π which may be held in one or morememories MEM’. Such data Π may be imbalanced. Therefore, indiscriminate use of such data in off-the-shelf training procedures may lead to low performance trained models. In in order to address this issue, a training facilitator system TFS (simply referred to herein as “facilitator”) is proposed herein. Such facilitator system TFS allows processing the training data in a particular way in order to address this data imbalance challenge, in particular intra-class imbalance. Broadly, the training facilitator system TFS computes certain auxiliary data, referred to as masks, which can be used in the training. Training may be framed as a numerical procedure to find a solution to an optimization (problem). The optimization is driven by an objective function F. That, is, the training is framed to adjust parameters of the model, so as to improve the objective function. The masks may used in the training to control or direct operation of the objective function F. The masks will allow more focused consideration of information in the training data as will be detailed further below. However, before explaining more particulars of the training facilitator system TFS envisaged herein, and in order to assist such explanations, reference is first made first Figure 2. This Figure illustrates general principles of spatial data acquisition as may be relevant for segmentation operations. Broadly, the imaging apparatus IA, or any measurement system configured for the acquisition of spatial data, may include a signal source SS and, co-operative therewith, a data acquisition unit DAQ. In data acquisition, signal source SS is energized to issueforth an interrogating signal ^. This interacts with the system from which one wishes togather information, such as a patient PT. Upon such interaction, a response signal ^ isissued from the interrogated system PT. The response signal ^ is then acquired bydata acquisition unit DAQ. The data acquisition unit DAQ may include a sensor arrangement suitably arranged to be sensitive to the responsive signal ^. The DAQ may be capable of natively registering in digital form. However, this is not a requirement herein. An analogue detectionarrangement may be used instead, such as X-ray sensitive film in a sealed cassettehousing, which may then be post-digitized. In either case, but not necessarily so, the response signal may be registered as spatially resolved intensities, or other measurement values. The so registered values form spatial data e, or even spatio-temporal data. The registered response ^ may be converted by an A / D conversion ofthe data acquisition unit DAQ into a numerical representation to represent such valuesthat form the spatial data. The data itself may be in two, three or higher datadimensions. One dimensional data is not excluded herein. However, the following will mostly appertain to 2D, 3D or even 4D dimensional data, such as projection imagery, an individual tomographic slice of a tomographic volume, or indeed the volume itself(3D), or a time series thereof (4D). The medical imagery e provided as themeasurement data may be of the tomographic type, although projection imagery is not excluded herein. Thus, any reference herein to “image”, “image data”, or similar, is a reference to spatial data in projection domain or image domain (post reconstruction).Outside the medical field, the spatial data / imagery e may include geologicalmeasurements such as in exploration, seismic, hydrology measurements or any other. Sonar and / or radar measurement data is also envisaged herein in engineering embodiments. The acquired data may be amplified. filtered, de-noised or otherwise cleansed orconditioned, as needed in the specific application domain. The acquired data data maybe represented in matrix form by a set of numbers or as tensors of higher dimension (3D or higher), as needed. Thus, spatial data as used herein represents data comprises individually addressable data elements (pixels, voxels) that correlate to spatial positions in 3D or 2D space. Thus, each pixel / voxel (we shall simply use theterm “pixel” in the following as a convenient shorthand), corresponds to a particularlocation 3D, in particular on or within the interrogated system, such as patient PT. Theso acquired imagery e may then be forwarded to the data consumer DC and processedtherein as mentioned earlier or as will be more detailed below. The interrogated system PT, such as patient PT, may be understood to be made up ofa number of physical objects E or structures of interest spatially arranged insurroundings R. For example, such physical objects E may include organs, parts of organs, groups of organs, tissue types or groups of such tissues, etc. What constitutessuch physical objects of interest E and what constitutes their spatial surroundings Rwill be a function of the imaging purpose. For example, if the purpose of medicalimaging is screening for nodules in a chest X-ray or mammogram, the physical objects of interest E are such nodules, with the surrounding tissue considered surrounding structures R of lesser interest for the task envisaged. Depending on the imaging taskat hand, in some embodiments, the physical objects E of interest may be of differenttypes, and / or the surroundings may include structures of different types. Thus, as sketched at the foot of Figure 2, the acquisition process α may beconceptualized as a mapping E → e from physical objects or phenomena E andsurrounding(s) R to the image data e. Thus, as such, the (raw) image data e soacquired represents both, physical objects of interest E and their surrounding(s) R.In more detail, the physical objects of interest E may be represented in the acquiredimagery e as objects of interest ^, whereas the surroundings R may be representedas background information ^. Thus, the imagery e may be thought to be made up ofobjects of one or more types, usually they are plural objects ^j of plural types, andbackground ^. The objects ^ = ^^ and the background ^’ are defined as (sub)sets inthe set theoretical sense of pixels that make up the image data (eg a volume or 2Dimage). Thus, set theoretic operations such the taking of unions " ∪” and intersections" ∩” are applicable, and may be indeed used herein, as will be explored later below inconnection with training based on certain types of objective functions F.Reference is now made to Figure 3 which illustrates the segmentation operation ^ asmay be envisaged herein in different “flavors”. Specifically, Figure 3A shows the imagedata as such, with objects ^ = ^^, and background ^’. Segmentation allows singlingout in automated fashion those image objects ^ of different types that arerepresentative of the physical objects E=Ej of interest and of different types or of a single type, as the case may be. Figure 3C illustrates a sophisticated form of segmentation, instance segmentation. In such instance segmentation it is not only theobject type that is identified, but each such instance ^^, ^^, ^^ (only 3 instances areshown) are identified by the segmentation individually rather than collectively. A collective standard segmentation is illustrated in Figure 3B. Segmentation can be understood as the identification / classification of the set of image pixels that make up the image objects of interest, either collectively or individually as shown in Figures 3B, C respectively. Identification may include identifying the location and extent (size) ofthe sets. In general, segmentation operation ^ of image e may be formulized as:^: E → ^=(^1,.., ^j, … ^N), such that the fibres of segmentation operator ^ (“segmentor”)is a partition of the image e ∈ ^ so segmented: ^ = if i≠j wherein ^^^is the pre-image of a given class cj.Because the image objects are sets, certain set theoretic concepts, such as union andintersection (overlap) can be readily applied thereto, and will be done so herein below. Reference is now made to Figure 4 which shows more details of the mentioned machine learning training facilitator system TFS in the context of a machine learning arrangement MA.The training system TS uses an objective function F to train model M based on trainingdata ^ as may be held in a memory MEM. Once the model M has been so trained inone or more training cycles, it can be used by the data consumer DC as described above, in particular for segmentation tasks by image processor IP, etc. Training may be a one-off, or may be done repeatedly in multiple cycles, using, in each cycle, the model as trained in a previous cycle as an initial configuration, and so on. Models are structures that are processable by training algorithms in computational manipulation. The ML model M is made up of an architecture, and parameters that are adjusted during the training to meet an objective. That objective is represented by the objective function F. It is the objective function F that measures whether the objective has been met. The output of training is hence a set of such parameters for the model. Said differently, the output of training is the trained model having those parameters as determined at the conclusion of the training. The training may proceed in iterations. The architecture of the model and some of its features are hyper-parameters that are chosen once, but are not adjusted during training. Such hyper-parameters are the number of layers in a neural network type model, as indeed may be used herein in embodiments. Other hyperparameters may pertain to features of the objective function F itself.Prior to processing by the training system TS of the training data ^, the trainingfacilitator system processes the training data Π into certain auxiliary data which can beused in the training. In particular, such auxiliary data may include masks that are applied to the training imagery to so control and direct the operation of the training system TS. In particular, the masks allow determining which information from the training data the objective function is to take into account. Thus, the training may be made more “focused” to address the above mentioned data imbalance problem, in particular intra-class imbalance. Operation of the training facilitator system TFS is now explained in yet more detail,with reference to the foot of Figure 4. The training data ^ is received at input port INand processed by a mask generator MG as will be described in more detail below. Themask generator MG generates a set of masks m(Π) for the training data Π. The masksm are output at output interface OUT. The output masks m may then be used inconjunction with the objective function F by the training system TS. Preferably, thetraining data Π is pre-processed by a pre-processor PP into pre-segmented data, andthis is so received at input port IN and processed into masks as described herein below. Thus, this pre-processing may be done by an instance segmenter component which may be automatic, or may be done instead by human-guided annotation in a purely manual or semi-automated manner. In order to distinguish the segmentation operation σ for which the model M is trained from such pre-segmentation operation, the pre-processor PP’s operation may be referred to herein as an annotation operation. For example, the pre-processor PP may be arranged as a viewing station that allows a human expert to review training image data e’, such as raw medical imagery, of historical examinations, as may be retrieved from a medical storage, and annotate same using annotation tool(s). Such annotation tool may include a computer systemthat is capable of visualizing the historic imagery ^^ ∈ ^ on a display device. In additionthe computer system CS may support a pointer tool for example that allows the userto annotate or delineate instances or examples of objects ^ of interest, such as imagestructures, that the expert believes the model to be trained should recognize. For example, if the model M is to be trained for segmentation in mammography, the model annotated objects may pertain to instances of nodules, such as a calcification or others. Thus, a single training image e’ may include the same class of objects (eg, those thatrepresent such nodules), but different instances thereof. Thus, the annotationsawarded by expert are such that that they are capable of coding for both, class and instance: Thus, the annotation data may be conceptualized as a double index (c,k), c- > class and k-> index. As the pointer tool moves from one object to the other, the instance index is incremented, thereby classifying each pixel into instance of classes(c,k). The instance index k may be a whole number that is incremented k+1->k witheach annotation event for a separate object, thus yielding a list of objects, 1st, 2nd, 3rd,on so forth. The order in which the objects appear in this list is immaterial. Thus, the input imagery is assumed herein to already include some segmentedinformation that preferably captures individual instances of image objects ^ of interestThe annotations (c,k), awarded per pixel or region per training image e’, and done whichever way, include not only spatial information c, but also instance information k, Such double information may be formalized by the above introduced double indexingscheme (c,k), one index (c ) pertaining to spatial information in the image domainwhere the raw image data e’ is thought to be situated, and the other index k that represents the particular instance of this image object. Not all training images e’ in the training data Π include annotations of objects of interest, as some such training images will serve as negative examples. But of the training images e’ that do include such annotated instances of objects ^ of interest, will also but include background^^^^′. In general, the area of background will be largerthan that of the annotated objects ^′: there is more information on background thanthere is on actual objects of interest ^′. This imbalance is addressed by the masks m generated herein by generator MG. The masks are configured herein for “judicious” suppression of background in favor of objects. Thanks to the masks, the model can be trained to recognize instances of the same class, thus enabling in good approximation instance-based segmentation. Thus, the masked data and their use by objectivefunction F appears to “incentive” the training system TS for instance basedsegmentation behavior it has been observed. The above described imbalance may be termed as one between foreground and background. In addition, or instead of the foreground vs background imbalance, and of particular benefit herein, is that the proposed mask generation is also capable of addressing the intra-class imbalance:The masks generated herein afford improved consideration in training the sizeimbalance within each, or some, foreground class.The training imagery e’ in training data ^ are of similar type than the image data ^ thatis processable later by the fully trained model during deployment as mentioned abovein Figures 1, 2. Indeed, the training data ^ is preferably so sourced or is so generatedsynthetically that its statistical properties, its distribution, etc, its collective properties roughly correspond to the statistical properties of the population whose imagery is expected to be seen by the model at the clinical site where deployment is intended.However, as such, the training data ^ is in general different to the one used or seen indeployment. Thus, as has been done above and will be done below, the “primed”notation e’ will be used to denote training image specimens e’ ∈ Π, in distinction todeployment image data e to be segmented post training. A similar primed notation willbe used and has been used above to denote objects ^’ and background ^ ^^^′ in suchimagery e’. The training data Π may include chest imagery e’ of prior examinations from the same or different patients done in the past. As such they are different from deployment data e’ as may be processed post training. Thus, the training data Π may include historical or synthetically generated imagery e’ of the type of interest, such asmammograms or CT volumes or others, and the proposed facilitator system FTSproduces masks m(e’) for such corpus of imagery {e’} in training data.The general concept of masks is illustrated in Figure 5, to which reference is nowmade. In contrast to the underlying image e’, the associated mask m(e’) is a datastructure distinct from the underlying image but may have the same domain size anddimension as the underlying image e’ does. The mask m is a function operable on theimage e’. It may be viewed as a scalar field, or indicator function, operable on image data m: e-> s, with “s” a set of codes. The codes determine the manner or processing of the masked image by objective function F in training (on which more further below). For example, if the volume e’, such a CT or MRI volume, of three dimensions n x m x ^, then so is the mask m of E’ that is produced by the training facilitator system TFS. A mask, such as a binary mask, may include / define an exclusion portion XP and an admission portion AP. The admission portion is illustrated as a “hole”. The two portions describe or define associated exclusion area XA and admission area AA in the imagery e’ for which the mask was prepared. The admission area AA generally includes the image information of the in-image object of interest, and it is only such information from the total image e’ that is admitted for consideration, in particular by the objective function F, during training by training system TS. Any other pixels are in the exclusion area XA, thus are excluded from consideration by the function F in training. If m is mask, so is its “inversion”, where the former exclusion portion XP is the new admission portion, and vice versa, former admission portion AP is the new exclusion portion. Inthe inverted mask, the exclusion portion XP is now surrounded by admitted portion AP.Thus, the admitted portion may not necessarily be conceptualized as á “hole” as illustrated in the Figure, but in many cases this may be indeed so.The mask m may be understood as a scalar function that assigns to each pixel valuev in the imagery a certain code s(v), and that code determines whether or not thespecific pixel is to be admitted or not into consideration by the training system in particular function thereof, in particular the objective function used by the training algorithm. Thus, the mask may be considered an n-dimensional indicator function where the coding may be binary {0,1}, and can be applied to the image values of imagee’, with n being the dimensions of the image e’. Thus, such a mask m(e’) defined forthe respective training image e’ may confer a “masking” of the objective function by extension:-Fm := F ° m(e’) (1),where “°" indicates functional composition / concatenation.However, below we will drop the superscript to denote but the function F and its masking with the same symbol, to so disencumber notation. Thus, when training system TS applies objective function to training imagery e’, thetraining system may cause evaluation mask m(e’) first, before considering theunderlying pixels. The above was explained with reference to a “hard” mask, a binary mask that is. However, this is not at the exclusion of soft masks as opposed to such hard masks.Soft masks assign a range of values rather than just binaries {0,1} to the pixels thatmade up the underlying imagery e’. Such soft mask entries can be considered asweights that respectively down- or upweight the extent to which the image informationis to be considered. For example, the masked-out image pixels in the exclusion area XA are to be considered to a lesser extent than the pixels that are in the admission area AA (thus, are not masked out). A relative weighing scheme may be used, or it is either XP pixels that are down weighted, or it is admission area AA pixels that are up- weighted, as needed.If a pixel v(x,y) in an image e is within the admission area, mask m instructs objectivefunction F to use the value of the pixel at the given in-image position (x,y). Thus, withinthe admission area, the mask acts like the identity function. Outside the admission area, mask m instructs function F to ignore the pixel value, by assigning an agreed code s, eg s =0, which is then interpreted by function F accordingly. Thus, function Fmay be implemented as callable routine with a logic (such an if-then instruction),capable of interpreting the codes it receives from the mask. One way to implement such mask is by pixel-wise multiplication. The admission portion comprises unity values 1’s, whilst the exclusion portion comprises 0’s. Multiplying such mask withimage e yields zeros in the exclusion area, and whilst the pixel values within admissionzone are preserved by multiplication with unity, and can be passed on to objectivefunction F for processing. in this example, the “zeros” are the codes s that determinethe function F to ignore the corresponding pixel value as per a pre-arranged logic operation. A similar pixel-wise multiplication operator can be used for soft masks.An (intrinsic) size of mask m may be considered as the size, in pixel area, of theadmission area AA. Thus, the intrinsic size of the mask is different from the earlier mentioned domain size, which corresponds to the size of the imagery to which the mask is to be applied. In general, unless otherwise stated herein, a reference to “size of mask” or similar, is to be construed as a reference to the said intrinsic size. Reference is now made to Figure 6 which shows the mask generation operation of mask generator MG in more detail. Generally, the mask generator processes an annotated or otherwise pre-segmented corpus of training spatial data (eg, imagery) and masks same to obtain training data of object-instance-oriented character for used in a supervised ML training setup TS. Figure 6A) shows schematically an annotated image e’, such as a tomographic image slice made of pixels as provided by the pre-processor PP. As mentioned above, such pre-processor PP may be implemented by a human expert operating computing system that supports an image annotation tool, such as contour delineator, a CADfunction, or any other Such function capable of labelling / annotating training imagery e’.Three objects of interest are shown as annotated, representative of lesions. Asexample illustration, three objects ^′ ^^^ objects are represented as three dark circles,a large one and two small ones, marked up with as annotations A1-3. The annotatedObjects ^′ ^^^ are surrounded by background ^^′, illustrated in black. The annotatedobjects ^′ ^^^ may represent different instances of the same class. The imagery e’ soannotated with annotations A1-3 for objects said objects ^′ ^^^ may be understood asground truth data g. Fig 6A represents a single example of such ground truth data, but in general they may be a plurality of such ground truth images g=gj, each based on different underlying training images e’=e’j. The following mask generation process will be illustrated for this given example of ground truth data g, with the understanding that the same processing may be applied to some or each ground truth data specimen gj.The number and appearance of annotated objects ^′ ^^^in the example image e’ ofFig 6A is purely for illustration and in no way limits the general principles described herein.The mask generator MG may process the ground truth data gj∈ Π. Specifically, maskgenerator MG is operable to process the ground truth data g comprised in the trainingdata Π. Fig 6B illustrates examples of masks as may be output by generator MG as a result of this processing.Four examples of such generated masks m are shown at a)-d) in Figure 6B forillustration: three “pure” masks m(g^j) illustrated at b)-d), and one “mixed” or “combo”mask m(g^′∗). In Fig 6, the illustrations represent the respective mask applied to the underlying image e’. Each pure mask m(g^′1-3) is dedicated to a particular instance of object of interest, such as m(g^′j) for object ^′j,. Thus, there is one single mask per object instance g^′1- 3. And in each such pure mask, a single instance is featured. In Fig 6B, b)-d), the respective hatchings represent the respective exclusion area XA, whilst the no-hatched area, the “hole”, illustrates the respective admission area AA. As can be seen, in each masked image, the respective admission area includes a single instance of theannotated object instance ^’j, plus a margin u of surrounding background ^^ , shown asa black ring-like structure. The admission portion AP is thus smaller than the originalimage e’ as shown in Figure 6A, has smaller size r. The margin u, or thickness of thebackground area / volume ^^ , representing part of the admission area, may be constantthroughout the different mask instances m(g^′j), although the size r may differaccordingly with object ^’j size. By including a small fraction of the background allows for better definition of the object contour. And by keeping the thickness of the background portion “(ring”) constant across the masks, in essence encourages the training algorithm to ignore any background contribution of the per mask loss. Thus, it is predominantly, if not only, object-of-interest information that is considered in the per mask, and hence per object, loss, as will be explored later, in more detail below at eqs (3)(4).The illustrations and explanations above and below of the masks m are in terms ofadmission area AA and exclusion area XA. Thus, the exclusion portion XP and admission portion AP of the respective mask is defined accordingly by suitable codess to result in the above explained ring-shaped admission area AA, including both, asingle respective object instance ^’j and a fraction f (subset) of the surroundingbackground area ^^. The addition of the fraction f (subset) of the surroundingbackground area ^^ allows for better definition of the object instance’s contour, but withmore weight given to the object itself at the expense of the background information. In the pure masks, their admission areas include only background and a single such instance. However, it may be possible that in some cases the background also includes portions of neighboring object instances. However, no second object is wholly included, but, if at all, only partly so. This may happen if the objects are very tightlypacked. Thus, in some cases, the background in the admission area AA may be“tainted” by one or more fragments of such neighboring other objects. Thus, it is only a single instance that is included as a whole in the respective admission areas of the masks as generated herein. Thus, such masks may not exclude the presence of fragments of such neighboring object instances. Preferably however, there are no such fragments in some or each mask. One way to implement such “ringed-masks” is based on the respective object’s contour as defined by annotation, and as supplied by pre-processor PP. The contour may be scaled up to so expand or enlarge the admission portion / area to the pre-set margin size u. Thus, the mask generator MG may include an expansion module EXP to accomplish. In other embodiments the expansion module EXP may be implementedby using a process called "dilation". This is an image processing operation where pixelsin reference location, such as the center of a structure (e.g. 3x3 square, or otherneighborhood patch), take on the values of their neighboring pixels according to apredefined pattern in that neighborhood patch, thus, in effect, expanding (i.e. dilating)the contour. However, using such contour expansion may be merely one manner of operation envisaged herein or the mask generator MG. For example, in alternative embodiments,the ringed-mask may be obtained by splines, or other.In addition to such pure ringed-masks, which are dedicated to one single instance of object of interest, there is a combo mask, shown to the very left as mask m(g^∗′).Compared to the “pure” masks m(g^^′) as shown at b)-d) of Figure 6B, the combomask is m(g^′) is much enlarged and includes in its admission area essentially and preferably all of the background of the raw data e’ of Figure 6A, in addition to one singleof one the objects, such as ^^′ or another, as illustrated at a). Thus, in general, giventhe number of instances N, the system MG produces N+1 instances of such masks / masked ground truth data g^, one for the combo mask and the remaining N forthe different object instances. This can be repeated for each ground truth image gj, toobtain one mask per occurrence of object of interest ^’. The purpose of the combo-mask is to add information on the background, largely separate from the object instances. Thus, the background may still be processed by the loss function (as will be discussed below in more detail), but in “turn”, separate from the object instances. The inclusion of the single instance object (any can be used) allows for better definition of the background.The above operation at Figure 6 has been illustrated for the ground truth part of thetraining data. However, the training data Π includes, in addition to the ground truthdata, the tentative predictions M(e’j) of the model M, given a certain model θ parameter,produced in the training.The prediction will be denoted herein as pj. Thus, each pj is a prediction of the givenmodel Mθ at a given parameters θ applied to a given training image ej’ ∈ Π. Thus, thetraining data Π comprises the ground truth’s gj, their masks m(gj), and the predictionsMθ (ej) = pj, with e = {ej} specimens / examples of training imagery. Because eachtraining image ej’ has a mask m(ej) as illustrated, there is a natural association betweenthe ground truths and their pj’s (pj, m(ej)=m(^^^) =: (pj, gj). As the model Mθ is to betrained for segmentation operation ^, the predicted segmentation pj = Mθ(e’j) yields predicted objects, and the above described mask generation at Figs 6A, B can be applied equally to the pj’s, to so yield the ringed masks as shown in b)-d) of Fig 6B. Such ringed masks for the prediction objects will be simply referred to herein as m(pj). Thus, the above masking scheme with mask size for each instant of predicted object, with a background margin around it yields pairs of associated image subsets (m(pj),gj) that can be used for a supervised training setup as administered by training system TS, as will be one explored in more detail. In case of the predictions pj, no human- provided annotation are needed, as the predicted objects can be readily obtained by the class labels as retuned by predicted segmentation Mθ(e’j).The above referred objective function F may be used with training data (m(pj ),gj) intraining by the training system TS. In more detail, the training system TS implements a training algorithm. This training algorithm may be driven by the objective function F. The objective function F may formulated as loss / cost function or a utility function. For simplicity we will be referring to the objective function as loss function F, or simpler still as “loss F” throughout, with the understanding that mutatis mutandis, a utility function setup may be used instead. The training may be formulated as an optimization task, in particular as minimization of loss F, whilst the dual formulation may be a maximization of utility.In general, the loss F is operable to compare ground truth with predictions per instance,and summing over such instances to form the total loss that is to be improved during training, in one or more iterations using a suitable numerical algorithm to adjust parameters θ of parameterized model Mθ:- argmin θ F =β ∑^ ^[ ^^(^^) , ^(^^ )] = −^ ∑ ^ ^^ ^[ ^^(^ (^ ^)) , ^(^^)] (2)Thus, the total loss F is made up of partial losses, each partial loss representing theindividual loss incurred by each instance k considered in the training. The training may proceed in iterations to adjust model parameters multiple times over, until stopping condition is fulfilled, such as when the loss drops under a threshold, or after a preset number of iterations, etc. In the second expression to the right of eq(2), some of the training imagery e’kmay be summed over repeatedly, depending on the number ofobject instances. For example, in the illustration in Fig 6, the training image e’ at Fig6A) will appear at least 3 times, because there are three objects represented. Because of the combo image Fig 6B), a), the image e’ may hence feature 4 times in different summand in (2). Consistent with the earlier introduced notion, summation in (2) is overobject instances k, with D a distance measure, ^^(^^) the masked predictions, and thegk’s the masked per instance ground truths (Fig 6B). The summation runs up to N+1,because of N object instances k, and the (optional, but preferred) one combo maskedimage.The loss F is a real-valued functions that maps image-subsets into a number.Specifically, as envisaged herein, the loss F compares using a suitable distancemeasure D[.,.], sub-sets of training imagery ej’ of spatial data, such as is defined by(m(pj),gj)). Some such cost functions compare the ratio of overlap (intersection ∩ )between the two subsets versus the two sets’ union ∪. Some such loss functions F isof the Dice loss-type, which can be formalized as follows, wherein each summand in (2) has the form of:- with one of the summands pertaining to the combo-masked image (Fig 6B, a) ) and associated prediction p*, just as is done for the pure masks. In case Dice loss D, a minus sign is appended so that the optimization (2) may be formulated as one of minimizing the loss. However, this is optional as a dual formulation in terms of anmaximation of Dice as is (without the minus sign) may also be contemplated.The size of the subsets may be measured by a respective suitable area size, or the respective number of counts of pixel encompassed by the respective subset. As can be seen in (2), a constant weight factor β<1 may be used for all summands. This weight factor may be chosen as β = 1 / (N+1). Thus, thanks to the proposed masking, each summand (partial / per instance loss) may be weighted the same. There is no establish an elaborate unequal weighting scheme as may have been used in the past for other Dice type losses. Thus, using the example of Fig, the loss may be written ^ as − ^∑^^^^ ^^^^ (^^, ^^) , with “g1” the combo-masked image.As an alternative to (3), certain Dice loss cognates may be used, such as lossreweighting, as described by Szu-Yeu Hu et al in their paper „Multimodal Volume-Aware Detection and Segmentation for Brain Metastases Radiosurgery“, published 15 August 2019, arxiv:1908.05418.Thus, as indicated at eqs (2),(3), the processing by such cost / loss function F mayproceed per instance k, and masks so produced, and their sum is considered in thecost function F when optimizing the sum. The parameters θ is adapted so as to improvethe objective function such as through a reduced loss value (a number) as returned by the Dice loss function. The number of parameters may be very large. The notation θ is used herein for a set of parameters that are adjusted θ = { θ1, … θL}, where whole number L may be large: it may run into the thousands or larger still, depending on the type and complexity of the ML model at hand. For example, of the model M is of the neural network (“NN”)-type, in particular of the convolutional neural network (“CNN”)- type, the said parameter may include filter coefficients of convolutional filters that may be used in various layers of which the CNN is made up.In Figure 6B, the lower row at b)-d) illustrates pure masks, as the admission areasthere included only background, and a single such instance. However, it may bepossible that in some cases the background also includes portions of neighboring object instances. However, no second object is wholly included, but, if at all, only partly so. This may happen if the objects are very tightly packed. Thus, in some cases, the background in the admission area AA may be “tainted” by one or more fragments of such neighboring other objects. Thus, it is only a single instance that is included as a whole in the respective admission areas of the masks as generated herein. Thus, such masks may not exclude the presence of fragments of such neighboring object instances. Preferably however, there are no such fragments in some or each mask. Reference is now made to the flow chart of Figure 7 which shows steps of computer implemented methods of providing such mask, and method that may use the above described mask development scheme and its uses downstream, in training and deployment / inference or testing. Broadly, flow chart Figure 7A pertains to the fashioning of the above described instance focused masks, flow chart Figure 7B relates to the use of such masks alongside training data in method a training a machine learning model for segmentation; and the flow chart in Figure 7C pertains to method of deployment / inference or testing of a machine learning model trained on such masked data. Referring first to the mask construction computer-implemented method in Figure 7A,this may be understood as a method of facilitating the downstream training of MLmodel. In particular, masks are generated in the method in order to address the imbalanced training data set problem. At step S710 an existing corpus of training data is accessed. The training data may include spatial data. In particular, training data may include imagery such as medical imagery of two, three or higher dimensions as needed. At step S720 ground truth data items are constructed from the training data set. This may include annotations by human experts, or may include synthetically generating such data, or may include any other way of defining ground truth data. Definition of such ground truth data may include contoured instances of in-image objects, that is, subsets of image pixels. Thus, the proposed set-up is mainly envisaged for a supervised learning scheme. Thus, the training data herein is assumed to have its training image data items, such as individual images, pre-annotated or pre-segmented. Each such annotated image item from the training imagery (a projection image or a tomographic slice, or indeed the whole or part of a tomographic volume) includes annotated one or more object instances of a given class, or one or more instances of objects of interest of different classes. The so object instance based (pre-)annotated or pre-segmented training data items constitute ground truth data to be later used in the training. Include definition of groundtruth. The pre-annotation for one or more object instances may be called prior-masks.Such prior-mask are usually tightly defined and follow the contours of the in-image object instance. Such pre-annotated object instances may represent different lesions, spatially separated. The so prepared ground truth data with its prior-masks is then received at step S730. At step S740, individual object instance focused masks are generated as explainedabove, with an admission area that includes the whole of the given object instance andin addition includes a margin of background around it. The masks’ admission areas AA (illustrated at b)-d) of Figure 6B) have thus a ring-shaped structure, and may thus be referred to herein as ringed-masks. In addition to this, the combo-mask as illustrated above at the very left (at a) ) of Figure 6B is generated, which includes an admission area focused on preferably the whole of the background and in addition thereto, to one single instance of an object. The ringed-mask may be generated by expansion (eg scaling up) of the prior-mask’s admission areas, as such tightly defining the objects contours, but now expanded to include the ring of background around the object. Geometrical routine tools may beused to implement this. For example, the above mentioned spline curve(s) may be fitthe annotated contours as per the prior mask. The area so circumscribed by the spline curve, may then be scaled up. Alternatively, a smallest shape primitive (such assquare / cube) that includes the spline curve area is fit and this is then expanded.Any of the objects processed in the pure masks (see b)-d) of Figure 6B) may be used to fashion the combo mask. The (optional, but preferred) combo mask together with the pure masks so produced may be made available at step 750 for training or other forms of processing. Whether to use such a combo mask may depend on the choiceof the loss function or objective function more generally: a pure background mask maybe used instead of a combo mask in some cases. Use of the combo mask may bebeneficial, if the objective function is based on Dice loss, to so foster better modellearning of properties of the background. Such ringed-masks are also computed for the predictions pk= Mθ(e’k), with e’kthe training data item (image) from which the related ground truth mask was generated. Thus, pairs of masked data (m(pk),m(gk)) are made available.Turning now to the training method as per Figure 7B, at step 810 N+1 masked data(m(pk),m(gk)) is received, including N pure masks and optionally one combo mask,.and masks for the respective predictions are received. N represents the number ofindividual instances of objects of interest. such as lesions, or other.At step 820 the masked data (m(pk),m(gk)) used in the training. In particular, the so masked data controls the operation of the loss function used by the training scheme. Thus, the loss function measures, for each instance, the amount of correspondence between m(pk) and m(gk). In particular, the overlap of these subsets are measure, preferably normalized. The loss function may use a Dice-type loss measure D. Such comparison will yield a per mask cost. However, because the masks considered herein are focused on individual object instances k, the losses are per instance. The total loss will then be the sum of all such per mask loss. The model parameters may then be adapted in one or more iterations to improve loss, until a stopping condition is met. Once the training has concluded after a suitable number of iterations, at step S830 the trained model M is then output given its currently adjusted parameters. In the above, at training step S830, any suitable training scheme may be used, such as gradient based, in particular back-propagation based methodology, as needed. The model’s architecture may be any one of a neural network, decision trees, or support vector machines SVM, as needed. In particular, if neural networks (“NN”), CNNs may be preferred on account of the use of spatial data. In CNNs, convolutional filters coefficients are among the learnable parameters. They are thus adjusted in the training. ´ For classification tasks, such as the segmentation task of main interest herein, the NN model may be made up of a run of convolutional hidden layers, followed downstreamby one or more fully connected layers, with a softmax-layer as the output layer wherethe feature maps from previous layer are combined into a classification result. This result may be output as a vector, where a given entry represents a score, probabilityor measure for an associated class. However, purely convolutional models, with nofully connected layer(s) may be considered herein in alternative embodiments. Insemantic segmentation, a softmax-function may be applied for some or every pixel, to obtain a vector of probability / scores per pixel. The part of the algorithm that includes adjusting the parameters may be driven by any suitable numerical scheme such as gradient based as said, such as conjugate gradients or stochastic gradients, Nelder-Mead, as needed and appropriate in the circumstance at hand. Turning now to the flow chart in Figure 7C, the so trained model may be accessed or may be otherwise be made available at step S910. At step S920 the so trained model is then applied to deployment data such as may befound in everyday clinical practice, such as to a mammogram or any other image data.The deployment or test data is different from the above-mentioned training data. The model may then be output at step S930 an estimate for a segmentation, such as for a lesion that may be suspect. Alternatively, the model is used for testing. The trained model’s output may be made available. It can be stored, displayed or processed, as the case may be to thereby suitably inform diagnosis, treatment, planning for surgery and for radiotherapy, as needed. Segmentation tasks considered herein include binary segmentation and multi-class segmentation. The proposed set-up can be used with classical segmentation that does not necessarily distinguish object instances, but can be used also for instance-based segmentation. Indeed, with the improved masking proposed herein, and the equal weighing as may be used in the cost function, the processing may allow approximating in good measure instance-based segmentation which is preferred in some clinical applications where distinguishing between different instances of lesions may be preferable. It should be noted that embodiments of the invention are described with reference to different subject matters. In particular, some embodiments are described with reference to method type claims whereas other embodiments are described with reference to the device type claims. However, a person skilled in the art will gather from the above and the following description that, unless otherwise notified, in addition to any combination of features belonging to one type of subject matter also any combination between features relating to different subject matters is considered to be disclosed with this application. However, all features can be combined providing synergetic effects that are more than the simple summation of the features. While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. The invention is not limited to the disclosed embodiments. Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing a claimed invention, from a study of thedrawings, the disclosure, and the dependent claims.In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude a plurality. A single processor or other unit may fulfill the functions of several items re-cited in the claims. The mere fact that certain measures are re-cited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope.

Claims

Brainlab AG Attorney’s File: B19045WO / AO CLAIMS1. A system (TFS) for processing training data for facilitating training of a machinelearning model (M) in a segmentation task, comprising: a) an input interface (IN) through which is receivable training data (e), suchtraining data comprising spatial data and capable of representing i) instances ofin-image objects and ii) background; b) a mask generator (MG) capable of generating masks (m) for the trainingdata, wherein a given such mask pertains to a single such instance of the saidobjects, referred to herein as an object instance, and further pertains to only asubset of the background surrounding the said object instance, andc) an output interface (OUT) for providing the so generated masks to facilitate the training of the machine learning model.

2. System of claim 1, wherein the mask generator (MG) is capable to so generatethe given mask based on a respective prior mask for the said object instance,wherein such prior mask pertains only to the said object instance, at theexclusion of background, and wherein the generating of the mask by generator (MG) is based on expanding the prior mask to then include the said subset ofbackground.

3. System of claim 1 or 2, wherein the mask generator (MG) is capable to sogenerate a further mask, that pertains to the, or another, object instance, andfurther pertains to the whole of the background or to a subset thereof that islarger than the surrounding background subset.

4. The system of any one of the preceding claims, wherein the said object instanceis representative of a clinical feature.

5. The system any one of the preceding claims, wherein the training data includesmedical imagery.

6. A training system (TS) for training a machine learning model based on themasks provided by the training data provider system (TFS) as per any one ofthe preceding claims.

7. The training system of claim 6, driven by an objective function wherein the saidobjective function is configured for per mask processing of the training data.

8. A machine learning based system (MS) including the machine learning model(M) as trained by the system of any one of claims 6-7.

9. The system of claim 8, wherein the system is any one of medical, orsurveillance.

10. The system of claim 8 or 9, capable of instance segmentation.

11. A computer-implemented method for processing training data for training amachine learning model in a segmentation task, comprising the following steps:i) receiving (S730) training data (e), such training data comprising spatial data and capable of representing i) instances of in-image objects and ii) background; ii) generating (S740) masks (m) for the training data, wherein a given such maskpertains to a single such instance of the said objects, referred to herein as an object instance, and further pertains to only a subset of the background surrounding the said object instance, and iii) providing (S750) the so generated masks to facilitate the of training of the machine learning model.

12. A method for training a machine learning model in a segmentation task, basedon training data processable by the method of claim 11.

13. A method of inference or testing, based on a machine learning model trained ontraining data processable by the method of claim 11.

14. Using masks for training data in training a machine learning model, the saidmasks obtainable by processing as per method of claim 11.

15. A program which, when running on a computer or when loaded onto a computer,causes the computer to perform the method steps of the method according to any one of the preceding claims; and / or a program storage medium on which the program is stored; and / or a computer comprising at least one processor and a memoryand / or the program storage medium, wherein the program is running on thecomputer or loaded into the memory of the computer; and / or a signal wave or a digital signal wave, carrying information which represents the program; and / or a data stream which is representative of the program.

Images