Method for predicting a tissue property of an anatomical structure

Abstract

The present invention relates to a method for predicting a quantitative image volume representing a tissue property of one or more anatomical structures. The method comprises: obtaining corresponding quantitative and qualitative image volumes of a first set comprising one or more complete or incomplete anatomical structures of a plurality of subjects; segmenting the quantitative and qualitative image volumes to obtain corresponding quantitative and qualitative mask volumes of a second set forming digital 3D representations of the one or more anatomical structures; training an artificial intelligence network by using the corresponding quantitative and qualitative mask volumes as processed or unprocessed of the second set; and inputting at least a qualitative image volume as processed or unprocessed to the trained artificial intelligence network to predict a quantitative image volume comprising predicted image voxel values of the one or more anatomical structures of a subject, the qualitative image volume comprising the anatomical structure(s) of the subject.

Description

[0001] METHOD FOR PREDICTING A TISSUE PROPERTY OF AN ANATOMICAL

[0002] STRUCTURE

[0003] FIELD OF THE INVENTION

[0004] The present invention relates to a method for predicting voxel-wise tissue properties, such as fat fraction or tissue density, from a qualitative volumetric image dataset, such as a T1-weighted magnetic resonance imaging (MRI) dataset, of an anatomical structure, which may be for instance, a muscle, a bone or an organ. The present invention also relates to an apparatus for implementing the proposed method.

[0005] BACKGROUND OF THE INVENTION

[0006] Quantitative analysis of tissue is an important prerequisite for medical diagnosis and prognosis analysis. The interpretation of medical images, such as MRI images, is largely qualitative with anatomical structures and tissue types differentiated based on their relative intensities. The anatomical information derived from qualitative MRI images is often sufficient for diagnosis without the need for quantitative tissue measurements. However, in certain instances, quantification of specific tissue properties is important for optimisation of patient care. For example, the percentage of fat in muscle tissue or the quality of tendon tissue, are important prognostic factors for rotator cuff repair surgery, while bone density is a critical consideration for implant positing. While MRI can theoretically be used to determine any physical, chemical or physiological parameter, such quantitative analysis requires the acquisition of two or more image volumes and thus increases acquisition times, limiting their clinical applicability. Thus, such images are often not clinically available for analysis.

[0007] For example, while quantitative MRI fat-suppressed MRI sequences are available to assess voxel-wise fat fraction (allowing for the volumetric mapping of fat fraction values), these sequences are not required for diagnosis and are therefore often usually not acquired for diagnosis of musculoskeletal conditions. Other image modalities, such as computed tomography (CT), also quantify specific tissue properties, for example tissue density, but may not be applicable due to the associated negative effects, such as irradiation. Alternatively, tissue properties can be assessed using biopsy. However, such measurements are invasive and analysis hyper-localised.

[0008] Furthermore, an analysis of the complete anatomical structure may also not be possible if it is not contained within the image field of view. To date no method exists to predict the tissue properties of a portion of an anatomical structure outside of the image field of view, as is also required, for example, to calculate volumetric averages, such as the whole muscle average fat fraction, or to analyse the spatial distribution of a tissue property.

[0009] Without availability of quantitative image data, qualitative or unreliable and inaccurate quantitative analyses of tissue properties are often performed on qualitative images. For example, fat fraction of the rotator cuff muscles is qualitatively characterised according to the criteria defined by Goutallier on a single T1-weighted MRI slice, or tissue contained within the segmented muscle volume from a volumetric T1 -weighted MRI image is characterised as 100% fat or 100% muscle in each voxel using tissue differentiation methods, such as intensity thresholding or artificial intelligence-based (Al- based) segmentation techniques.

[0010] BRIEF DESCRIPTION OF THE INVENTION

[0011] To enable accurate quantification of tissue properties without the need for quantitative image data, the proposed computer-implemented method and system may be used to quantitatively map a tissue property of an anatomical structure from a qualitative image volume in which the complete volume of interest of an anatomical structure may or may not be visible. The proposed method is thus a computer- implemented method of data acquisition or data processing. Thus, the method or at least some of the method steps are implemented in software. Thus, at least some of the method steps can be considered as computer-implemented steps.

[0012] According to a first aspect of the invention, there is provided a method for predicting a quantitative image volume representing one or more anatomical structures as recited in claim 1.

[0013] The proposed method addresses and resolves the existing drawbacks in anatomical tissue analysis from qualitative image data, which may include image data with a restricted field of view. By integrating advanced artificial intelligence (Al) techniques, the proposed approach can accurately quantify the value of a specific tissue property for each voxel contained within the anatomical structure using only qualitative images of the anatomical structure.

[0014] This reduces the need for additional non-standard imaging, thereby reducing image acquisition time and costs and in the case of imaging using ionising radiation, irradiation dose. These predictive models optionally overcome the limitations of restricted field of view of certain imaging modalities and the inaccuracy of tissue property analysis in qualitative images. As a non-limiting example, the proposed method enables the prediction of a volumetric map of fat fraction of a complete muscle, such as the supraspinatus, from standard qualitative shoulder MRI sequences, such as T1-weighted MRI sequences of the shoulder, which do not typically include the entire muscle.

[0015] According to another non-limiting example, the proposed method can be used to determine the density of a bone, such as the humeral head, as required to plan the position of implants or anchors from qualitative MRI, such as a T1 -weighted MRI. For accurate bone density analysis, currently an additional CT image would be required. This method would enable quantitative bone density mapping without the need for additional irradiating imaging.

[0016] According to a second aspect of the present invention, there is provided a computer program product comprising instructions for implementing the steps of the method when loaded and run on a computing apparatus.

[0017] According to a third aspect of the invention, there is provided a data processing apparatus, which is configured to carry out the method according to the first aspect of the present invention.

[0018] Other aspects of the invention are recited in the dependent claims attached hereto.

[0019] BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The invention will now be described in more detail with reference to the attached drawings, in which:

[0021] • Figure 1 shows a block diagram schematically illustrating an example data processing system where the teachings of the present invention may be applied;

[0022] • Figure 2 is a flowchart illustrating the training phase of the proposed method according to a first embodiment of the present invention;

[0023] Figure 3 is a flowchart illustrating the application or inference phase of the proposed method according to the first embodiment of the present invention;

[0024] Figure 4 schematically illustrates the process of obtaining a complete image volume starting from an incomplete image volume; Figure 5 summarises the processes described in the flowcharts of Figures 2 and 3;

[0025] • Figure 6 summarises the application phase of the proposed method according to the first embodiment of the present invention;

[0026] • Figure 7 summarises the application phase of the proposed method according to a second embodiment of the present invention; and

[0027] • Figure 8 summarises the application phase of the proposed method according to a third embodiment of the present invention.

[0028] DETAILED DESCRIPTION OF THE INVENTION

[0029] Embodiments of the present invention will now be described in detail with reference to the attached figures. It should be noted that the figures are provided merely as an aid to understanding the principles underlying the invention and should not be taken as limiting the scope of protection sought. Where the same reference numbers are used in different figures, these are intended to indicate similar or corresponding features. As utilised herein, “and / or” means any one or more of the items in the list joined by “and / or”. As an example, “x and / or y” means any element of the three-element set {(x), (y), (x, y)}. In other words, “x and / or y” means “one or both of x and y.” As another example, “x, y, and / or z” means any element of the seven-element set {(x), (y), (z), (x, y), (x, z), (y, z), (x, y, z)}. In other words, “x, y and / or z” means “one or more of x, y, and z.” Furthermore, the term “comprise” is used herein as an open-ended term. This means that the object encompasses all the elements listed, but may also include additional, unnamed elements. Thus, the word “comprise” is interpreted by the broader meaning “include”, “contain” or “comprehend”. As used herein, unless otherwise specified, the use of the ordinal adjectives ’’first”, ’’second”, “third”, etc. to describe a common object, merely indicate that different instances of like or different objects are being referred to, and are not intended to imply that the objects so described must be in a given sequence, either temporally, spatially, in ranking, or in any other manner.

[0030] Figure 1 schematically illustrates an example imaging, data acquisition, or data processing system 1 , which is configured to carry out the proposed method as described later in more detail. The system comprises a data processing unit 2, which can access a memory unit 3 configured to store a set of medical images of an anatomical region of a subject. The system also comprises one or more machine learning networks 4, 5, 6 or pieces of Al software (i.e., an Al application), which may for instance be artificial neural networks, such as convolutional neural networks. In the present example and as explained later in more detail, the system 1 comprises a first Al network 4 configured to segment anatomical structures, and also referred to as a segmentation Al network, a second Al network 5 configured to predict missing portions of anatomical structures, and also referred to as a shape or structure prediction Al network, and a third Al network 6 configured to predict quantitative image volumes representing one or more anatomical structures starting from qualitative image volumes, and also referred to as an image volume prediction Al network. These Al networks may further include sub-networks to carry out specific tasks. The system in this case further comprises a user interface 7 to receive user inputs from an operator, and a data output unit 8, which may comprise a display and / or means to output digital information, such as an input / output port.

[0031] The first embodiment of the present invention is next explained in more detail with reference to the flowchart of Figures 2 and 3. It is to be noted some of the steps may be optional and / or some of the steps may be carried out in a different order. The flowchart of Figure 2 illustrates the training phase, and Figure 3 illustrates the application or execution phase of the proposed method. At step 11 , medical image data, which may be a series or set of images (including single images), are acquired or accessed. The images may be acquired using a medical imaging system, such as an MRI or CT system, or the data may be accessed from a database, such as a medical image archive. The images in the raw image dataset may be 2D (single slice of a 3D image) and / or 3D images. The image data consist of a set of corresponding quantitative image volumes and qualitative image volumes of one or more anatomical structures from a plurality of training subjects. In other words, the image data comprise pairs of corresponding quantitative and qualitative image volumes of one or more anatomical structures from a plurality of subjects. A quantitative image volume and a qualitative image volume correspond to each other, if they are taken of the same structure (i.e. one or more anatomical structures of the same individual). In the case that the anatomical structure is deformable (e.g. organ or muscle), these corresponding or paired quantitative and qualitative image volumes would be acquired with the structure in the same configuration such as for example, during the same examination. The image volumes are comprised of one or more image slices composed of a plurality of image voxels defined by a location and a value, also referred to as an intensity value. In common MRI sequences, the measured signal intensity reflects a combination of the spin lattice (T1 ) and spin-spin (T2) relaxation times, and the proton density of a tissue, resulting in a qualitative image. In contrast, the voxel intensity of a quantitative image or parametric map, represents a tissue (or anatomical structure more broadly) characteristic or property in an anatomically defined location. The quantitative image of an anatomical structure therefore represents a tissue property by the voxel intensity values, e.g. tissue density, as represented by the Hounsfield units in a CT image, or fat fraction as directly represented in a fat fraction MRI image as calculated from water only and fat only MRI images. The quantitative images would contain the complete region of interest of an anatomical structure. The qualitative MRI image of the same (yet possibly incomplete) structure would not directly or indirectly represent the tissue property with its voxel intensity values alone.

[0032] Data preprocessing is performed at step 12 to generate corrected or pre-processed 3D volumes. Images taken during clinical workflow may suffer from large variabilities in imaging protocols, especially the resolution and field of view. These variations may be addressed at pre-processing step 12, containing image interpolation and resizing the entire image to fit a single predetermined size.

[0033] The corrected or pre-processed 3D volumes containing one or more anatomical structures, which in this example are one or more muscles, are segmented at step 13 to obtain segmented or labelled volumes of the one or more anatomical structures, i.e., the segmented or labelled volumes show the anatomical structure(s). Volume segmentation is a process used to partition or label a volume into multiple regions or segments (in this case into two segments so that a first segment is identified with a first bit value, while a second, different segment is identified with a second, different bit value), making it easier to analyse and interpret specific structures within the volume. More specifically, at step 13, the 3D volumes are segmented by labelling the one or more anatomical structures with at least a first label, and remaining elements (i.e., the background) in the 3D volume with at least a second, different label to obtain mask volumes or labelled or segmented image volumes, which are digital 3D representations of the one or more anatomical structures. In this example, the first label has a bit value of 1 , and the second label has a bit value of 0, or vice versa. Thus, in the present example, the obtained digital 3D representations are binary mask volumes, which are 3D volumes highlighting the respective anatomical structure with a given bit value. However, it would be possible to have more than one anatomical structure inside a mask volume, where a respective anatomical structure is represented with its individual (integer) value. In this example, the first label has a first voxel value, the second label has a second, different voxel value, and so on. The following segmentation types are available:

[0034] • Manual segmentation involves a radiologist or technician manually outlining regions of interest. While accurate, it is time-consuming and subject to inter- and intra-observer variability.

[0035] • Semi-automatic segmentation combines manual input with automated processes to improve efficiency and consistency.

[0036] • Automatic segmentation: uses algorithms to automatically identify and segment regions. This method aims to be both time-efficient and reproducible.

[0037] In the present example, the segmentation is carried out by the first Al network 4, which in this example is a deep convolutional neural network (DCNN). Alternatively, the segmentation of the specific anatomical structure could also be done manually or semiautomatically. In some configurations, to validate performance of the automatic segmentation, comparison models may be used.

[0038] Referring to Figure 4, if the full region of interest of a structure is not contained within the qualitative image, the second Al network 5 can be trained to predict the missing volume. The corresponding complete binary mask volumes from the quantitative image volumes and the incomplete binary mask volumes from the qualitative image volumes can be used to train the second Al network 5, which may be a convolutional neural network, to predict the volume of the missing portion of the anatomical structure(s) of interest. A 2D representation is illustrated in Figure 4 for simplicity.

[0039] The upper right-hand side of Figure 4 shows the complete region of interest of an anatomical structure contained within the quantitative image volume, while the upper lefthand side of Figure 4 shows the incomplete region of interest contained within the qualitative image. At step 14, the binary mask volumes of the qualitative image volumes can be increased towards the direction of the missing part of the structure(s) of interest using voxel padding with the value of the background or with other values (excluding in this case the label value used to identify the anatomical structure(s)) to allow for a prediction outside of the original size of the anatomical structure(s). If the region of interest is contained within both image datasets, the qualitative images are cropped to the size of the quantitative image volumes, or alternatively the quantitative images are cropped to the size of the qualitative image volumes. At step 15, data augmentation is applied to the binary mask volumes, i.e. the resized quantitative and qualitative binary mask volumes to create a first training data set. Data augmentation is in this case used to artificially generate additional new data from existing data for the purpose of training the second Al network 5.

[0040] At step 16, pairs of corresponding complete and incomplete binary mask volumes can then be used to train the second Al network (e.g. a DCNN) 5 as shown in the upper part of Figure 4 to allow prediction of complete qualitative binary mask volumes. The accuracy of the second Al network 5 can be verified using additional paired quantitative and qualitative image data.

[0041] At step 17, the padded incomplete binary mask volumes of the qualitative image volumes are applied to the trained second Al network to predict the complete binary mask volumes, including the missing volume(s) of the anatomical structure(s) as shown in the lower part of Figure 4. In other words, this step involves predicting the complete binary mask volumes, including the missing volume(s) of the anatomical structure(s) of interest, from the incomplete binary masks of the qualitative image volumes. As is further shown in the lower part of Figure 4 illustrating the application phase, the input incomplete binary mask volume is in this example first increased in size by using voxel padding before the incomplete binary mask volume is fed to the second Al network 5.

[0042] At step 18, all the binary mask volumes, i.e. both the quantitative and qualitative binary mask volumes, are then applied to their corresponding image volumes to obtain masked image volumes, i.e. masked quantitative and qualitative image volumes, by for example, taking the Hadamard product of the respective binary mask volume and its corresponding image volume, which is the image volume used to derive the respective binary mask volume. In the resulting masked image volumes, voxels inside the anatomical structure(s) of interest maintain their original voxel values (obtained from the corresponding original image volumes) while voxels outside of the segmented structure(s) of interest are set to zero or to another value characterising the background. At step 19, data augmentation is applied to the masked image volumes, i.e. both the masked quantitative and qualitative image volumes, to create a second training data set.

[0043] At step 20, with the second training data set, the third Al network 6 is trained to later predict a masked quantitative image volume from a qualitative image volume. In other words, pairs of corresponding masked quantitative image volumes and masked qualitative image volumes are used to train the third Al network (e.g. a DCNN) 6 to allow prediction of a masked quantitative image volume from a masked qualitative image volume during the application phase. For example, the network 6 could be trained to predict the fat fraction, represented as the value of each voxel contained within the region of interest. Alternatively, the network could be trained with paired CT (which are quantitative images) and qualitative MRI images to predict Hounsfield units from a qualitative MRI in the region of interest. Optionally, a secondary mapping of the quantitative data could be performed, for example, to categorise fat fraction or another property into discrete classes (e.g. 1 = 0-10%, 2 = 11-20%....) prior to network training. The third Al network 6 would then be trained to alternatively or in addition predict these secondary mapped values from the corresponding qualitative image. The accuracy of the third Al network 6 can be verified using additional paired quantitative and qualitative image data.

[0044] The application phase of the method is explained next in more detail with reference to the flowchart of Figure 3. At step 21 , a qualitative image volume of at least part of the anatomical region of analysis of a subject under examination is accessed. In this example, the accessed qualitative image volume is not among the image volumes used to train the third Al network 6. The qualitative image may be acquired using a medical imaging system, such as an MRI system, or the data may be accessed from a database, such as a medical image archive. The image in the raw image dataset may be a 2D (single slice of a 3D image) and / or a 3D image. Data pre-processing is performed at step 22 to generate a corrected or pre-processed 3D volume. Images taken during clinical workflow may suffer from large variabilities in imaging protocols, especially the resolution and field of view. These variations may be addressed at pre-processing step 22, containing image interpolation and resizing the entire qualitative image volume to fit a single predetermined size. Similarly to step 13, at step 23, the qualitative image volume is segmented to obtain a binary mask volume. This step follows the principles of step 13. Similarly to step 14, at step 24, the size of the qualitative image volume and its binary mask volume may be increased to include the full anatomical structure(s) of interest. At step 25, the increased binary mask volume is input to the trained second Al network 5, which was trained at step 16, to predict a complete binary mask volume of the anatomical structure(s) of interest. It is to be noted that steps 24 and 25 are not needed, if the qualitative image volume acquired at step 21 already includes the entire region of interest. Similarly to step 18, at step 26, the (predicted) binary mask volume is applied to the image volume acquired at step 21 to create a masked qualitative image volume by performing a mathematical operation between the qualitative binary mask volume and the qualitative image volume as explained above in connection with step 18. At step 27, the masked qualitative image volume is input to the third Al network 6, which was trained at step 20, to predict a masked quantitative image volume. In this example case, the resulting masked quantitative image volume is a volumetric map of fat fraction of the anatomical structure(s). However, if the voxel values of the masked quantitative image volume do not directly represent the tissue property of interest, additional operations, i.e. postprocessing, could be applied to the masked quantitative image volume, for example, the multiplication of each voxel value by a calibration factor or the discrete classification of each voxel value to obtain the volumetric tissue property map. Additional postprocessing may also be applied to the masked quantitative image volume such as up- or down- sampling. Thus, at step 28, the predicted masked quantitative image volume is optionally postprocessed to generate a volumetric tissue property or quality map of the region of interest.

[0045] Figure 5 schematically illustrates both the training phase and the application phase of the proposed method according to the first embodiment. In this example, the qualitative image volumes and the quantitative image volumes are segmented by their respective segmentation model 4. The flowchart of Figure 6 summarises the application phase of the proposed method according to the first embodiment of the present invention. This example also includes the optional size increase of the binary mask volume.

[0046] The processes described herein may be executed by a computing system, such as the one shown in Figure 1 , or by another processor. Artificial intelligence (Al) techniques can be integrated into the software within the system 1 and accessed by the processor 2 to 1) segment an image volume, 2) optionally predict the missing volume of an anatomical structure of interest in the form of a (binary) mask volume and 3) predict quantitative voxel values of an anatomical region of interest from a qualitative image volume of a subject under examination. Different network architectures may be employed. For example, Al models of at least one of the first, second and third Al networks could utilise a 3D U-Net structure according to the teachings of publication entitled “3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation” by Ozgun iqek et al. designed for segmentation, incorporating an encoder-decoder framework with multiple resolution levels for 3D images. The encoder may include convolutional, batch normalisation, and activation layers, followed by pooling layers. The decoder may include deconvolution layers and similar convolutional blocks, with skip connections linking encoder and decoder layers. The final layer adjusts feature dimensions to match the number of anatomical structures to be segmented, followed by a voxel-wise classifier. Implementation of the respective network may leverage a neural network framework, such as TensorFlow. Training and testing may be conducted on graphics processing units. During training, weights may be initialised randomly from a Gaussian distribution and updated using an adaptive moment estimation (Adam) optimiser for gradient descent, with a predefined initial learning rate. Extensive data augmentation may be applied during training, such as random rotation, flipping, resizing, shearing to allow for better generalisation of the network. The loss function may include voxel-wise mean squared error (MSE), smooth L1 loss, binary cross-entropy (BCE) with logits, dice loss, or cross-entropy or a combination of multiple of these loss functions.

[0047] In some configurations, to validate performance of the trained networks during application, results of comparison models may be used. A comparison model may comprise subnetworks of a model or additional networks trained on different training data. For example, a comparison model may include a network which performs multiple segmentations of an anatomical structure at once. A comparison of predicted data can be used as an indicator of network output reliability. For example, if comparison networks predict similar results as the applied network, the output is considered to be more reliable than if the comparison networks outputs differ from the applied network predicted result.

[0048] According to the second embodiment, alternatively to the prior masking of the input volume before input to the third Al network 6, the complete binary mask volume generated by the second Al network 5 could be an additional input of the third Al network 6 as shown in Figure 7. In this case, the third Al network 6 would have two inputs: 1 ) the qualitative image volume and 2) the complete binary mask volume resulting from the segmentation of the qualitative image volume. Such a network would output a masked quantitative image volume for volumetric tissue property map generation. Furthermore, such a network would also be trained with a set of qualitative image volumes and their incomplete (or complete) qualitative binary mask volumes (segmented from the qualitative image volumes) and corresponding quantitative image volumes and their complete binary mask volumes (segmented from the quantitative image volumes).

[0049] According to the third embodiment, missing volume of the region of interest and the quantitative masked image volume could alternatively be predicted by a single network as represented in Figure 8. In this case, the network would have two inputs: 1) the qualitative image volume and 2) the incomplete binary mask volume resulting from the segmentation of the qualitative image volume. Such a network would also be trained with a set of qualitative image volumes and their incomplete (or complete) qualitative binary mask volumes (segmented from the qualitative image volumes) and corresponding quantitative image volumes and their complete binary mask volumes (segmented from the quantitative image volumes).

[0050] 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 being not limited to the disclosed embodiments. Other embodiments and variants are understood and can be achieved by those skilled in the art when carrying out the claimed invention, based on a study of the drawings, the disclosure and the appended claims. New embodiments may be obtained by combining any of the teachings above. 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. The mere fact that different features are recited in mutually different dependent claims does not indicate that a combination of these features cannot be advantageously used. Any reference signs in the claims should not be construed as limiting the scope of the invention.

Claims

CLAIMS1 . A method for predicting a quantitative image volume directly or indirectly representing a tissue property of one or more anatomical structures, the method comprising:• obtaining (11 ) pairs of corresponding quantitative image volumes and qualitative image volumes of a first set comprising one or more complete or incomplete anatomical structures of a plurality of subjects;• segmenting (13) the quantitative and qualitative image volumes of the first set by labelling the one or more complete or incomplete anatomical structures with a first label, and remaining elements in the one or more image volumes with at least a second, different label to obtain pairs of corresponding quantitative and qualitative mask volumes of a second set forming digital 3D representations of the one or more anatomical structures;• training (20) an image volume prediction artificial intelligence network (6) by using at least the corresponding quantitative and qualitative mask volumes as processed or unprocessed of the second set; and• inputting (27) at least a qualitative image volume as processed or unprocessed of a subject under examination to the trained image volume prediction artificial intelligence network (6) to predict a quantitative image volume comprising predicted image voxel values of the one or more complete or incomplete anatomical structures of the subject under examination, the qualitative image volume comprising the one or more complete or incomplete anatomical structures of the subject under examination.

2. The method according to claim 1 , wherein the method further comprises postprocessing (28) the predicted image voxel values to determine a volumetric map of a tissue property of the one or more complete or incomplete anatomical structures from the predicted quantitative image volume.

3. The method according to any one of the preceding claims, wherein the predicted quantitative image volume is a predicted masked quantitative image volume, wherein the one or more complete or incomplete anatomical structures are represented by the predicted image voxel values, and the background is represented by a set of background voxel values, different from the predicted image voxel values.

4. The method according to any one of the preceding claims, wherein the method further comprises increasing (14) the size of the qualitative mask volumes of the second set to match the size of the quantitative mask volumes of the second set.

5. The method according to any one of the preceding claims, wherein the method further comprises applying (15) data augmentation to the quantitative and qualitative mask volumes of the second set as processed or unprocessed to obtain a first training data set.

6. The method according to claim 5, wherein the method further comprises training (16) a structure prediction artificial intelligence network (5) with the first training data set, and applying (17) incomplete mask volumes of the qualitative image volumes of the second set to the trained structure prediction artificial intelligence network (5) to predict complete mask volumes of the qualitative image volumes, including missing volumes of the one or more incomplete anatomical structures comprised in the qualitative image volumes of the first set.

7. The method according to any one of the preceding claims, wherein the method further comprises applying (18) the quantitative and qualitative mask volumes of the second set as processed or unprocessed to their corresponding quantitative and qualitative image volumes of the first set to create quantitative and qualitative masked image volumes of a third set, wherein the one or more complete or incomplete anatomical structures are represented by image voxel values of the image volumes of the first set, and the background is represented by a set of background voxel values, different from the image voxel values.

8. The method according to claim 7, wherein the method further comprises applying (19) data augmentation to the quantitative and qualitative masked image volumes of the third set as processed or unprocessed to obtain a second training data set, and training (20) an image volume prediction artificial intelligence network (6) with the second training data set.

9. The method according to any one of the preceding claims, wherein prior to inputting the qualitative image volume of the subject under examination to the trained image volume prediction artificial intelligence network (6), the method further comprises processing (22, 23, 24, 25, 26) the qualitative image volume of the subject underexamination to create a masked qualitative image volume, wherein the one or more complete or incomplete anatomical structures are represented by image voxel values of the qualitative image volume of the subject under examination, and the background is represented by a set of background voxel values, different from the image voxel values.

10. The method according to claim 9, wherein the processing comprises segmenting (23) the qualitative image volume of the subject under examination by labelling the one or more complete or incomplete anatomical structures with a first label, and remaining elements in the one or more image volumes of the subject under examination with at least a second, different label to obtain a qualitative mask volume forming a digital 3D representation of the one or more anatomical structures; and applying (26) the qualitative mask volume as processed or unprocessed to the qualitative image volume of the subject under examination to create the masked qualitative image volume.

11. The method according to claim 10, wherein prior to creating the masked qualitative image volume, the method further comprises increasing (24) the size of the qualitative mask volume to include the complete anatomical structure of interest, and applying (25) the qualitative mask volume as resized to a structure prediction artificial intelligence network (5) to predict a complete qualitative mask volume, including a missing volume of the one or more incomplete anatomical structures comprised in the qualitative image volume of the subject under examination.

12. The method according to any one of the preceding claims, wherein the predicted quantitative image volume is predicted by the trained image volume prediction artificial intelligence network (6) by using as inputs: the qualitative image volume of the subject under examination, and a qualitative mask volume as processed or unprocessed derived from the qualitative image volume of the subject under examination by segmenting the qualitative image volume of the subject under examination and optionally by including one or more predicted missing volumes of the one or more incomplete anatomical structures comprised in the qualitative image of the subject under examination; or a masked qualitative image volume derived from the qualitative image volume of the subject under examination by segmenting the qualitative image volume of the subject under examination to thereby obtain a qualitative mask volume, and applying the qualitative mask volume as processed or unprocessed to the qualitative image volume ofthe subject under examination, wherein in the masked qualitative image volume, the one or more complete or incomplete anatomical structures are represented by image voxel values of the qualitative image volume of the subject under examination, and the background is represented by a set of background voxel values, different from the image voxel values.

13. The method according to any one of the preceding claims, wherein the mask volumes are binary mask volumes.

14. A computer program product comprising instructions for implementing the steps of the method according to any one of the preceding claims when loaded and run on a computing apparatus.

15. A data processing device for predicting a quantitative image volume directly or indirectly representing a tissue property of one or more anatomical structures, the device comprising the means for:• obtaining pairs of corresponding quantitative image volumes and qualitative image volumes of a first set comprising one or more complete or incomplete anatomical structures of a plurality of subjects;• segmenting the quantitative and qualitative image volumes of the first set by labelling the one or more complete or incomplete anatomical structures with a first label, and remaining elements in the one or more image volumes with at least a second, different label to obtain pairs of corresponding quantitative and qualitative mask volumes of a second set forming digital 3D representations of the one or more anatomical structures;• training an image volume prediction artificial intelligence network (6) by using at least the corresponding quantitative and qualitative mask volumes as processed or unprocessed of the second set; and• inputting at least a qualitative image volume as processed or unprocessed of a subject under examination to the trained image volume prediction artificial intelligence network (6) to predict a quantitative image volume comprising predicted image voxel values of the one or more complete or incomplete anatomical structures of the subject under examination, the qualitative image volume comprising the one or more complete or incomplete anatomical structures of the subject under examination.

Images