Anatomical encryption of patient images for artificial intelligence

Abstract

An apparatus (10) for generating a training set of anonymized images (40) for training an artificial intelligence (AI) component (42) from a plurality of images (11) of people. The apparatus includes at least one electronic processor (20) programmed to: spatially map the plurality of images of people to a reference image (30) to generate an image (32) in a common frame of reference; divide the image in the common frame of reference into P spatial regions (34) to generate P sets (36) of image patches corresponding to the P spatial regions; assemble a set (38) of training images in the common frame of reference by, for each training image in the common frame of reference, selecting an image patch from each of the P sets of image patches and assembling the selected image patches to the training image in the common frame of reference; and, process the training images in the common frame of reference to generate the training set of anonymized images, including applying a statistical inverse spatial mapping to the training images in the common frame of reference, wherein the statistical inverse spatial mapping is derived from a spatial mapping (33) of the plurality of images of people to the reference image.

Description

TECHNICAL FIELD

[0001] The following relates generally to image processing technology, medical image diagnostic analysis technology, patient anonymization technology, artificial intelligence (AI) technology, and related technology. BACKGROUND

[0002] In the analysis of medical images, AI is increasingly being used. For example, an artificial intelligence classifier can be used to detect lesions, classify images to determine whether they depict a certain medical condition, etc. Typically, an AI component is trained using a set of training images, which are often labeled with a "correct" classification by a clinical expert (to guide the training). For example, with a clinician labeling tumors, clinical images with and without tumors can be used as a training set. The AI component is then trained to maximize the accuracy of distinguishing between images with and without tumors. EP 3 188 058 Al describes a method and apparatus for anonymizing image data in medical images, in particular images of facial bones or heads.

[0003] However, a problem arises in that the training images can be considered private patient data. Even if the metadata associated with the images is anonymized, the images themselves can potentially be identified as being of a particular individual, and can contain information about the medical condition of the individual. A trained AI component can potentially embed the training images. As a result, applicable patient privacy regulations can prohibit the distribution of an AI component trained on clinical patient images. This can be overcome by obtaining patient consent to use the images in training, but it is difficult to build a sufficiently large and diverse training set in this way, and to maintain an auditable record of all associated patient consent documents. Another approach is to synthesize training images, for example using anatomical models and imaging physics models to produce synthetic training images, but synthetic images can differ from real clinical images, which can not be apparent to a human viewer, but can introduce systematic errors into the resulting trained AI component.

[0004] Disclosed below are some improvements to overcome these and other problems. SUMMARY

[0005] In one aspect, an apparatus for generating a training set of anonymized images to train an AI component from images of a plurality of people. The apparatus comprises at least one electronic processor programmed to: spatially map images of a plurality of people to reference images to generate images in a common reference frame; partition the images in the common reference frame into P spatial regions to generate P sets of image patches corresponding to the P spatial regions; assemble a set of training images in the common reference frame by, for each training image in the common reference frame, selecting an image patch from each of the P sets of image patches and assembling the selected image patches to the training image in the common reference frame; and process the training images in the common reference frame to generate a training set of anonymized images, including applying a statistical inverse spatial mapping to the training images in the common reference frame, wherein the statistical inverse spatial mapping is derived from the spatial mapping of images of the plurality of people to the reference images.

[0006] In another aspect, a non-transitory computer-readable medium stores instructions executable by at least one electronic processor to perform a method of generating a training set of anonymized images for training an AI component from images of a plurality of people. The method comprises: spatially mapping images of a plurality of people to reference images to generate images in a common reference frame; partitioning the images into P spatial regions to generate P sets of image patches corresponding to the P spatial regions; assembling a set of training images in the common reference frame by, for each training image in the common reference frame, selecting an image patch from each of the P sets of image patches and assembling the selected image patches to the training image in the common reference frame; and processing the training images in the common reference frame to generate a training set of anonymized images, including applying a statistical inverse spatial mapping to the training images in the common reference frame, wherein the statistical inverse spatial mapping is derived from the spatial mapping of images of the plurality of people to the reference images.

[0007] In another aspect, a method of generating a training set of anonymized images for training an AI component from images of a plurality of people. The method comprises: spatially mapping images of a plurality of people to reference images to generate images in a common reference frame; partitioning the images into P spatial regions to generate P sets of image patches corresponding to the P spatial regions; assembling a set of training images in the common reference frame by, for each training image in the common reference frame, selecting an image patch from each of the P sets of image patches and assembling the selected image patches to the training image in the common reference frame; and processing the training images in the common reference frame to generate a training set of anonymized images, including applying a statistical inverse spatial mapping to the training images in the common reference frame, wherein the statistical inverse spatial mapping is derived from the spatial mapping of images of the plurality of people to the reference images.

[0008] One advantage is that the training image dataset is generated from the patient images using only parts of the patient images.

[0009] Another advantage is that the training image dataset is generated from the patient images without the need to extract patient specific information from the images.

[0010] Another advantage is that the image content of the patient images is anonymized in an irreversible way.

[0011] Another advantage is that the image content of the patient images is anonymized before the images are used to train the AI component.

[0012] Another advantage is that a training image dataset is generated in which each training image comprises parts of multiple patient images from different patients.

[0013] The presented embodiments can provide none, one, two, more or all of the above-mentioned advantages, and / or can provide other advantages which will be apparent to those of ordinary skill in the art having the benefit of this disclosure after its review and understanding. BRIEF DESCRIPTION OF DRAWINGS

[0014] The present disclosure can be presented in terms of various components and components arrangements, as well as various steps and steps arrangements. The drawings are only for purposes of illustrating the preferred embodiments and should not be interpreted in limiting the present disclosure.

[0015] Figure 1 An illustrative apparatus for generating a training set of anonymized images for training an AI component from images of multiple people according to the present disclosure is schematically shown.

[0016] Figures 2-7 Examples of images generated by the apparatus of Figure 1 are shown. DETAILED DESCRIPTION

[0017] Systems and methods for anonymizing real clinical images (not just the associated metadata) are presented below. To this end, the following procedure is disclosed. First, a set of real clinical images {R} r＝1,…,R is mapped to a reference image. The reference image can be a clinical atlas image, or also a “typical” clinical image (optionally taken from the set {R}). The spatial mapping will typically be a non-rigid spatial registration, and will yield a spatial transformation Z r for each image r e {R}. The result is a set of spatial transformations {Z r} r＝1,...,R (or if one image is taken as the reference image, the result is {Z r} r＝1,...,R-1 ).

[0018] Next, the spatially mapped images are divided into a set of spatial regions {P} p＝1,…,P The set of spatial regions {P} can be defined using a grid of straight lines, or the spatial regions can be defined along anatomical lines. This provides R patches corresponding to each spatial region p of the set of spatial regions {P}.

[0019] Then, a set of training images {N} n＝1,…,N Each training image n e {N} is constructed by randomly selecting one of the R patches corresponding to each spatial region p, and then combining the selected patches into a training image n. The resulting N patches form an image in the reference image space, which is undesirable because it does not capture the true distribution of sizes and shapes of the anatomical structures being imaged. To address this, a randomly selected inverse transform is applied to the image formed by each patch. In one approach, the randomly selected inverse transform is randomly selected from the set of spatial transforms {Z r}. In another approach, the statistical distribution of the parameters of the spatial transforms of the set {Z r} is determined, and inverse transforms are generated using these distributions.

[0020] A potential problem with this approach is that the boundaries of the patches can be discontinuous in the training images. For example, a limb skeleton can exhibit an artificial "break" at the boundary between two adjacent spatial regions. Whether this is a problem can depend on the nature of the AI component being trained. (Obviously, if the AI is being trained to detect limb fractures, this will be a problem; but if the AI is being trained to detect minor injuries, the likelihood that this patch boundary will be mistaken by the AI for an injury can be low).

[0021] Two approaches to solving this problem are disclosed. The first embodiment includes a process of performing smoothing at the boundaries. This can be most easily done prior to applying the inverse transforms, because prior to the inverse transforms, the boundary locations are the same for each patched image. In the second embodiment, the set of spatial regions {P} is designed to avoid having spatial region boundaries cut across major anatomical boundaries. For example, the liver of a reference image can be divided into a subset of spatial regions that are all entirely within the liver; each lung can be divided into a subset of spatial regions that are all entirely within the lung; and so on. This embodiment allows spatial region boundaries to be avoided from crossing anatomical boundaries, and is feasible because the set of spatial regions {P} is only delineated once, and then the same set of spatial regions {P} is applied to each of the {R} clinical images to generate the patches.

[0022] In some embodiments disclosed herein, the performance of the AI component trained on the chunked training set {N} can be easily ranked by comparing its performance to that of a similar AI component trained on the original set of clinical images {R}. The latter similar AI component, which can contain personally identifiable image information, will be discarded after it is used as a benchmark, and the AI component trained on the chunked training set {N} will be distributed to customers.

[0023] The disclosed system and method can be applied not only to training of AI components to analyze medical images, but more generally to training of any AI image analysis component that will be trained on a set of training images that all have the same underlying layout (e.g., face images for training a facial recognition AI, or images of people for training an AI to classify attributes of the photographed person, or retinal scan images for training a retinal scanner to perform user identification, are other examples of cases where the disclosed method can be useful). Moreover, the method is applicable to two-dimensional or three-dimensional (i.e., volumetric) images. In the case of volumetric images, the set of spatial regions {P} will be defined over the volume, and the image chunks will be volumetric image chunks.

[0024] Reference is made to Figure 1Fig. 1 shows an illustrative apparatus 10 for generating a training set of anonymized training images 40 to train an Al component 42 from images 11 of multiple people. For example, the apparatus 10 can receive images 11 of multiple people from an image acquisition device (also referred to as an imaging device) 12. The imaging device 12 can acquire images 11 as medical images of anatomical structures. In this example, the image acquisition device 12 can be a magnetic resonance (MR) image acquisition device, a computed tomography (CT) image acquisition device; a positron emission tomography (PET) image acquisition device; a single photon emission computed tomography (SPECT) image acquisition device; an X-ray image acquisition device; an ultrasound (US) image acquisition device; or other modalities of medical imaging devices. Additionally or alternatively, the imaging device 12 can be a camera (in which case the images 11 are portrait images) or a retinal scanner (in which case the images are retinal images) or the like appropriate to the particular Al component 42 to be trained, e.g., (video, RBG, infrared, etc.). It will be appreciated that the illustrative imaging device 12 is representative, and more generally, the image set 11 can be acquired by multiple different imaging devices. For example, the images 11 can be collected from a database such as a picture archiving and communication system (PACS) or other medical image repository 28 that can serve a hospital, hospital network, or other entity. The images of the image set 11 all have the same basic layout. For example, in the case of clinical images, the images 11 are all the same imaging modality and the same anatomical portion (e.g., all head MRI images, all liver MRI images, all chest CT images, etc.), and all the same or similar orientation (e.g., all sagittal head MRI images, or all coronal head MRI images, etc.), and use the same or similar imaging conditions (e.g., all use the same contrast, or all use no contrast) or imaging conditions spanning the operating space of imaging conditions to be used with the Al component. As another example, for use in training a retinal scanner, the images 11 can all be retinal images. As another example, for use in training a facial recognition system, the images 11 can all be portrait facial images. Further, the images 11 should statistically represent a demographic range and medical condition or conditions to be processed by the Al component 42 to be trained. If there are known confounding medical conditions that can cause misdiagnosis (or other erroneous output) by the Al, then the image 11 set can include examples of such confounding conditions. For example, if the images 11 are mammogram images, and skin folds in mammogram images are expected to adversely affect operation of the Al component 42, then the image 11 set (here, mammogram images) can optionally include some mammogram image with skin folds to improve robustness of the ultimately trained Al component.

[0025] Figure 1An electronic processing device 18, such as a workstation computer, or more generally a computer, is also shown. Alternatively, the electronic processing device 18 can be embodied as a server computer or multiple server computers, for example interconnected to form a server cluster, cloud computing resources, etc. The workstation 18 includes typical components, such as an electronic processor 20 (e.g., a microprocessor), at least one user input device (e.g., a mouse, a keyboard, a trackball, etc.) 22, and a display device 24 (e.g., an LCD display, a plasma display, a cathode ray tube display, etc.). In some embodiments, the display device 24 can be a component separate from the workstation 18.

[0026] The electronic processor 20 is operatively connected with one or more non-transitory storage media 26. The non-transitory storage media 26 can include, by way of non-limiting example, one or more of: a magnetic disk, RAID, or other magnetic storage medium; a solid state drive, flash drive, electronically erasable read-only memory (EEROM) or other electronic memory; an optical disk or other optical storage; various combinations thereof; etc.; and can be, for example, network storage, an internal hard drive of the workstation 18, various combinations thereof, etc. It should be understood that any reference to a non-transitory medium or media 26 herein should be interpreted broadly to encompass a single medium or multiple media of the same or different type. Likewise, the electronic processor 20 can be embodied as a single electronic processor or two or more electronic processors. The non-transitory storage medium 26 stores instructions that are executable by the at least one electronic processor 20. The instructions include instructions to generate a graphical user interface (GUI) 27 for display on the display device 24.

[0027] The apparatus 10 also includes or is otherwise in operative communication with a database 28 that stores the images 11. The database 28 can be any suitable database, including a radiology information system (RIS) database, a picture archiving and communication system (PACS) database, an electronic medical record (EMR) database, etc. For example, the database 28 can be implemented by a server computer and a non-transitory medium 26. The workstation 18 can be used to access the stored images 11. It should also be understood that the images 11 can be acquired by multiple imaging devices as previously described, and not necessarily just by the representative one illustrated image acquisition device 12.

[0028] The apparatus 10 is configured as described above to perform a method or process 100 for generating a training set of anonymized images. The non-transitory storage medium 26 stores instructions that are readable and executable by the at least one electronic processor 20 to perform the disclosed operations, including performing the method or process 100 for generating a training set of anonymized images. In some examples, the method 100 can be performed at least in part by cloud processing.

[0029] AsFigure 1 The illustrated and referenced Figure 2-7 An illustrative embodiment of the imaging examination workflow visualization method 100 is schematically shown as a flowchart. Figures 2-7 An example of the output of the operation of the method 100 is shown. Figures 2-7 The images shown in are two-dimensional X-ray images of a chest, but it will be appreciated that the method 100 can be applied to any suitable imaging modality or anatomical region of a patient. Without loss of generality, it is assumed that the images 11 form a set {R}, i.e. that there are R images in the set of images 11. Although Figure 1 Not shown in is the metadata associated with the images 11 is stripped, or at least any personally identifiable information (PII) contained in the metadata associated with the images 11 is stripped. Also, without loss of generality, it is assumed that the generated set of anonymized training images 40 forms a set {N}, i.e. that there are N anonymized training images. In general, there is no relationship between R and N, only that the number R of images 11 should be large enough compared to the number N of anonymized training images 40 so that the blocks formed by dividing the R images 11 into spatial regions P are “well mixed” in the anonymized training images 40. (To illustrate this, if for example R = 5 images 11 and N = 100 training images 40 are to be generated, then at least some of the training images will likely include mostly blocks extracted from one of the five input images 11, which is undesirable. On the other hand, if R = 1000 and N = 20, then it is likely that no anonymized training image contains one or two blocks from any one of the 1000 source images 11.

[0030] At operation 102, the at least one electronic processor 20 is programmed to spatially map the plurality of images 11 to a reference image 30 to generate images 32 in a common reference frame. When the images 11 are medical images, the reference image 30 can be an anatomical atlas image of an anatomical structure. When the images 11 are portrait images, the reference image 30 can be a face atlas image. Alternatively, in either of these examples, the reference image 30 can be one of the plurality of images 11 (in Figure 1 This option is indicated schematically in ). In the latter case, the reference image 30 can be manually selected from the set of images 11 - for example, a “typical” image is preferably selected as the reference image. (Typically, in this case, it is appropriate that the image is of an average size anatomical structure, and preferably does not have any significant abnormalities).

[0031] Figure 2 An example of the operation 102 is shown. Four images 11 are acquired by the imaging device 12 and / or retrieved from the database 28. Each of the four images 11 shown is from a different patient (in Figure 2"top" of the set). Image 11 is mapped to reference image 30 (in this case, an atlas image, and in Figure 2 "middle" of the set). The mapping can be performed using any suitable mapping algorithm known in the art (such as one described in A. Franz et al., "Precise anatomy localization in CT data by an improved probabilistic tissue type atlas", SPIE Medical Imaging: Image Processing, Vol. 9784, 2016). Such a Franz method provides a non-linear mapping function that allows all images 11 to be mapped to a common atlas reference 30. After mapping, all images 32 in the common frame of reference are comparable, e.g. a location / pixel in one image is related to a corresponding anatomical location in all other images. These images 32 are shown in Figure 2 "bottom" of the set.

[0032] Mapping operation 102 also outputs a set of mappings or transforms 33. In one approach, the mapping of each image in the set of images 11 to reference image 30 is one of the set of mappings 33. This results in R mappings 33 (or possibly R-1 mappings if reference image 30 is one of the R images 11). In another approach, the set of mappings 33 is output as a multidimensional statistical distribution. For example, if mapping operation 102 employs a parametric mapping algorithm, the distribution of each parameter in the R mappings of the R images 11 to reference image 30 is represented by, for example, a Gaussian fit to the mean and standard deviation of the parameter distribution.

[0033] At operation 104, the at least one electronic processor 20 is programmed to divide the image 32 in the common frame of reference into a plurality (designated, without loss of generality, as P) of spatial regions 34 to generate P sets of image tiles 36 corresponding to the P spatial regions. In some embodiments, the P spatial regions 34 form a rectilinear grid, while in other embodiments the P spatial regions 34 are aligned with regions of interest in the reference image 30, e.g. anatomical boundaries between organs in the reference image 30. In this embodiment, e.g. when the image 11 is a medical image, the boundaries of the P spatial regions 34 do not cross anatomical boundaries of anatomical structures in the reference image 30. The set {P} of spatial regions 34 is predefined, either automatically (e.g. using a computer-generated rectilinear grid) or manually drawn, e.g. drawn on the reference image 30 using a contour-drawing GUI 27 as used in contour-drawing of organs for radiotherapy planning. A hybrid approach is also envisaged, in which boundaries of major organs are manually drawn using a contour-drawing GUI 27 to define coarse spatial regions aligned with the organs or other anatomical structures, and then each coarse spatial region is itself automatically divided into a computer-generated rectilinear grid, defining the final set of P spatial regions 34.

[0034] The set of spatial regions 34 is suitably chosen to ensure anonymity of the resulting anonymized training images 40. To this end, the spatial regions 34 should be chosen small enough that no single tile 36 is individually identifiable. In addition, the number of spatial regions (P) should be large enough that the probability that a majority or all of the randomly selected tiles making up a given training image will come from a single image of the set of images 11 is statistically negligible. In addition, it is useful for the number of images 11, R, to be larger (and preferably much larger) than the number of spatial regions P, again reducing the probability that a majority or all of the randomly selected tiles making up a given training image will come from a single image of the set of images 11.

[0035] Figure 3A 、 Figure 3B and Figure 4 Two examples of operation 104 are shown. As shown in Figs. 2A and 2B, the image 32 in the common frame of reference is divided into a plurality of spatial regions 34. Figure 3A and Figure 3B The image 32 in the common frame of reference is divided into a plurality of spatial regions 34. Figure 3A An example of an embodiment in which the spatial regions 34 form a rectilinear grid is shown (e.g. 4 spatial regions are shown, and have the same size), while Figure 3BAn example of an embodiment in which the spatial regions 34 are aligned with regions of interest in the reference images 30 is shown (e.g., only 3 spatial regions are shown). It should be noted that as used herein, a "spatial region" refers to a spatial delineation used to partition an image 11 into blocks. The term "block" as used herein then refers to a portion of a given partitioned image 11 that corresponds to a given spatial region. Since there are R images 11, the partitioning operation 104 will yield R blocks for each spatial region that corresponds to a set of P spatial regions 34. There will be P sets of blocks corresponding to the P spatial regions, each set of blocks comprising R blocks. As Figure 4 shown, the spatial regions 34 are applied to the images 32 in the common reference frame to generate blocks 36. Each block 36 can be identified as a tuple comprising (image identifier, block number). The image identifier is preferably not a personalizing identification. Rather, for example, the R images can simply be numbered 1,..., R, and the number assigned to an image is the image identifier. Each block 36 corresponding to a given spatial region 34 is anatomically comparable across the images 32 in the common reference frame.

[0036] At operation 106, the at least one electronic processor 20 is programmed to combine the set of training images 38 in the common reference frame (see also Figure 5 ). To this end, an image block 36 is selected from each of the P sets of image blocks for each training image 38 in the common reference frame. The selected image blocks 36 are combined into a training image 30 in the common reference frame according to the spatial layout of the set of spatial regions 34. In some examples, selecting an image block 36 from each of the P sets of image blocks comprises randomly or pseudo-randomly selecting an image block. As used herein, the term "pseudo-random" and similar nomenclature has its usual and ordinary meaning in the field of computer science and refers to a process that is technically deterministic but produces results that are statistically similar to a random process and the results produced are not repeated for multiple runs of the pseudo-random process. For example, a common pseudo-random value generator comprises a number sequence generator that produces a deterministic sequence of numbers that has similar statistics to a truly random sequence of numbers and is very long (possibly infinite). For each run, a seed is selected to determine the starting point in the sequence. For example, the seed can be selected based on a set of low bits of the computer clock value at the start of the run, which is essentially a random number for a high-speed computer processor running at gigahertz or higher. Other pseudo-random generators, such as Monte Carlo simulation methods, are also applicable. Moreover, the term "random" as used herein should be understood to include embodiments that employ a pseudo-random value generator.

[0037] Figure 5An example of operation 106 is shown. For each spatial region 34, a random image identifier is selected by, for example, a random number generator, and a block from the set of R blocks 36 corresponding to that spatial region with that image identifier is selected. This is repeated for each spatial region in the set of P spatial regions 34. Thus, the selected blocks 36 are in general different images from the set of images 11, and are combined using the spatial regions 34 (see Fig. 3) to generate a set of training images 38 in the common reference frame. Figure 4 The training images 38 shown in the common reference frame include blocks 36 from 4 different patients. In practice, the number of blocks P can be much larger than 4.

[0038] In the method described above, there is some possibility that two (or even more) blocks from a single image in the set of images 11 can be included in a single training image 38 in the common reference frame. This is unlikely to be a problem as long as the number of images R is much larger than the number of spatial regions P, and preferably also much larger than the number of training images 40 produced, N. However, if it is desired to ensure that no training image 38 in the common reference frame has more than one block from a single image 11, then it is possible to check whether the set of final blocks combined for each training image has a duplicate (i.e. two or more blocks from a single image 11). If such a duplicate is found, then the training image in the common reference frame is discarded. This approach requires that R > P, and preferably R » P should hold.

[0039] Operation 106 is repeated until the desired number N of training images 38 in the common reference frame have been combined. The training images 38 in the common reference frame are anonymous. However, they do not represent the statistical variation in size / shape of the people (or imaged anatomical parts thereof) in the set of images 11. This is a result of the mapping operation 102, which results in the training images 38 all being within the common reference frame of the reference image 30.

[0040] Accordingly, at operation 108, the at least one electronic processor 20 is programmed to process the training images 38 in the common reference frame to generate a training set of anonymized images 40 that are representative of the statistical variation in size / shape of the person (or the imaged anatomical portion thereof) in the set of images 11. To this end, a statistical inverse spatial mapping is applied to the training images 38 in the common reference frame. The statistical inverse spatial mapping is derived from the spatial mappings 33 of the images 11 of the plurality of persons to the reference image 30. In one example, the spatial mappings 33 of the images 11 to the reference image 30 are inverted to form a set of inverse spatial mappings from which the statistical inverse spatial mapping is randomly or pseudo-randomly selected. In another example, the statistical distribution of the parameters of the spatial mappings 33 of the images 11 is computed to form a set of inverse spatial mappings, and the statistical inverse spatial mapping is generated from these statistical distributions. Optionally, operation 108 can also include a smoothing operation that can be performed at the boundaries of the image patches 36 to further generate the training set of anonymized images 40.

[0041] Figure 6 An example of operation 108 is shown. By applying the inverse atlas mapping, the training images 38 in the common reference frame are remapped to the original patient image space of the images 11 (see Figure 2 ). One of the mappings computed at operation 102 is randomly selected, inverted, and applied to the training images 38 in the common reference image 30 space to generate the training set of anonymized images 40.

[0042] Figure 6 An example of one of the anonymized images 40 is also shown. However, the anonymized images 40 can not have perfectly matching edges of the patches 36 to each other. As Figure 6 shown, the upper right and lower right patches 36 introduce an offset that is visible at the boundary of the lung. This offset can be explained by variations in image intensity in the original images 11 and / or imperfect mapping algorithms. The resulting anonymized images 40 should be as close to a real patient scan as possible. Accordingly, both the image intensity and edge structure at the boundaries of the patches 36 are optionally repaired by using smoothing or other suitable image processing techniques. For example, intensity values can be corrected by adjusting the intensity average computed near the patch boundaries, and mismatched edges can be locally offset by a non-linear image transformation. Operations 106 and 108 can be repeated to generate a new database of anonymized images 40. The original images 11 can then no longer be reconstructed from the set of anonymized images 40.

[0043] The operations 104-108 are described above as being performed in the context of anatomical image correlation. However, these operations 104-108 can be performed in the context of functional image correlation. That is, blocks 36 from images with matching patterns of functional data can be combined to form training images 38. To this end, pattern recognition operations are performed on the set of images 11 to form sub-combinations of images with correlated functional data across blocks 36. Such sub-combinations of blocks 36 are used to form training images 38. The pattern recognition operations and sub-combinations completed prior to creation of the training images 38 (i.e., operation 108) form can be used in the anatomical images to improve block correlation, e.g., by patient / organ size, disease state, location of abnormality, or clinical impact, etc. This data can come from image analysis, or supplemented with patient data (e.g., from the patient’s clinical record).

[0044] In some examples, the PACS database 28 can receive images 11 from multiple imaging devices 12 that can be located at one or more medical facilities. Blocks 36 from these images can be stored with the sub-combination specification until sufficient image data meeting the subset criteria is available to create additional training images 38.

[0045] Referring back to Figure 1 At operation 110, the at least one electronic processor 20 is programmed to train the AI component 42 on the training set of anonymized images 40. The trained AI component 42 can be deployed in the workstation 18 for use as part of a computer-aided clinical diagnostic workstation (e.g., a radiology workstation used by a radiologist in interpreting an imaging study). Anatomical images 11 can be used to train the AI component 42 of a medical diagnostic device, while portrait images 11 can be used to train the AI component 42 of a facial recognition device (e.g., the AI is deployed in a facial recognition scanner for controlling access to a restricted area). As yet another example, if the images 11 are retinal images, the trained AI component 42 can be deployed in a retinal scanner for controlling access to a restricted area. These are just some illustrative applications.

[0046] In some examples, the operation 110 can include verifying the training of the AI component 42 on the training set of anonymized images 40. To this end, the AI component 42 is trained on the training set of images 40 to generate a trained artificial intelligence component. A separate instance of the AI component 42 is trained on the original images 11 to generate a reference trained AI component. The performance of the AI component 42 trained on the training set of anonymized images 40 is verified by comparing the performance of the trained AI component trained on the training set of anonymized images 40 to the performance of the reference trained AI component (i.e., trained on the original images 11).

[0047] In general, the number of spatial regions P can be increased to provide a higher degree of anonymization of the anonymized image 40. As previously mentioned, the set of spatial regions 34 can have various spatial geometries and layouts. For example, "jigsaw puzzle" shapes, patterns, decomposition tiling can be used to define the spatial regions 34. In particular, the spatial regions can be defined in the reference image 30 in such a way that the boundaries of the spatial regions 34 do not cut through anatomical edges, or at least do not cut through anatomical edges perpendicularly. Figure 7 An example of an anonymized image generated using such tiling is shown. This can still result in errors at the boundaries between spatial regions 34, but they are greatly reduced and can be more easily corrected by smoothing or other image processing.

[0048] In some embodiments, to further anonymize the original image 11, as previously mentioned, the mean and variance of all the mappings 33 can be computed. From these, a random inverse mapping can be generated, further increasing the degree of anonymization.

[0049] In other embodiments, when blocks 36 are merged, the image content or anatomical type (e.g. gender) can be taken into account, so that only similar patient blocks are merged together. In another example, a registration algorithm can be used to merge the blocks 36.

[0050] The present disclosure has been described with reference to preferred embodiments. Modifications and alterations can occur to others upon reading and understanding the preceding detailed description. It is intended that the exemplary embodiment be construed as including all such modifications and alterations insofar as they fall within the scope of the appended claims or the equivalents thereof.

Claims

1. An apparatus (10) for generating a training set of anonymous images (40) for training an artificial intelligence (AI) component (42) based on images (11) of multiple people, the apparatus comprising at least one electronic processor (20) programmed to: The image spaces of the multiple individuals are mapped to a reference image (30) to generate an image (32) in a common reference frame; The image in the common reference frame is divided into P spatial regions (34) to generate a set of P image blocks (36) corresponding to the P spatial regions; The set of training images (38) in the common reference frame is combined by selecting an image patch from each of the P image patch sets for each training image in the common reference frame and combining the selected image patch into the training image in the common reference frame; and The training set for generating anonymized images by processing the training images in the public reference frame includes applying a statistical inverse space mapping to the training images in the public reference frame, wherein... The statistical inverse spatial mapping is derived from the spatial mapping (33) from the images of the plurality of people to the reference image.

2. The apparatus (10) according to claim 1, wherein, The image (11) of the plurality of people is: - A medical image of the anatomical structure, and the reference image (30) is an anatomical atlas image of the anatomical structure; or - A portrait image, and the reference image (30) is a facial atlas image.

3. The apparatus (10) according to claim 1, wherein, The reference image (30) is one of the images (11) of the plurality of people.

4. The apparatus (10) according to any one of claims 1-3, wherein, The P spatial regions (34) form a linear grid.

5. The apparatus (10) according to any one of claims 1-3, wherein, The P spatial regions (34) are aligned with the region of interest in the reference image.

6. The apparatus (10) according to claim 4, wherein, The images (11) of the plurality of persons are medical images of anatomical structures, and the boundaries of the P spatial regions (34) do not cross the anatomical boundaries of the anatomical structures in the reference image (30).

7. The apparatus (10) according to any one of claims 1-3, wherein, Selecting an image block (36) from each of the P sets of image blocks includes randomly or pseudo-randomly selecting an image block from each of the P sets of image blocks.

8. The apparatus (10) according to claim 7, wherein, Selecting an image block (36) from each of the P sets of image blocks includes randomly or pseudo-randomly selecting an image block from each of the P sets of image blocks having a matching functional data pattern.

9. The apparatus (10) according to any one of claims 1-3, wherein: The at least one electronic processor (20) is also programmed to invert the spatial mapping of the images (11) of the plurality of people to the reference image (30) to form a set of inverse spatial mappings, and The statistical inverse space mapping is selected randomly or pseudo-randomly from the set of inverse space mappings.

10. The apparatus (10) according to any one of claims 1-3, wherein: The at least one electronic processor (20) is also programmed to calculate the statistical distribution of parameters of the spatial mapping (33) from the images of the plurality of people to the reference image (30), to form a set of inverse spatial mappings; and The statistical inverse spatial mapping is generated based on the statistical distribution of the parameters of the spatial mapping from the images of the plurality of people to the reference image.

11. An apparatus for training medical diagnostic equipment, the apparatus comprising: At least one electronic processor (20) according to claim 1, programmed to generate anonymous images (40) from images (11) of a plurality of people, wherein, The images of the aforementioned individuals are medical images of anatomical structures; The at least one electronic processor is also programmed to train the artificial intelligence (AI) component (42) of the medical diagnostic device on the training set of anonymous images.

12. An apparatus for training a facial recognition device, the apparatus comprising: At least one electronic processor (20) according to claim 1, programmed to generate anonymous images (40) from images (11) of a plurality of people, wherein, The images of the aforementioned individuals are portrait images; The at least one electronic processor is also programmed to train the artificial intelligence (AI) component (42) of the facial recognition device on the training set of anonymous images.

13. An apparatus for training and validating a trained artificial intelligence (AI) component (42), the apparatus comprising At least one electronic processor (20) according to claim 1, programmed to generate anonymous images (40) from images (11) of a plurality of people; in, The at least one electronic processor is also programmed to: The artificial intelligence (AI) component is trained on the training set of anonymous images to generate a trained artificial intelligence (AI) component; The artificial intelligence (AI) component is trained on images of multiple individuals to generate a reference trained AI component; and The artificial intelligence (AI) component is validated by comparing the performance of the trained AI component with that of a reference trained AI component.

14. A method (100) for generating a training set of anonymous images (40) for training an artificial intelligence (AI) component (42) based on images (11) of multiple people, the method comprising: The image spaces of the multiple individuals are mapped to a reference image (30) to generate an image (32) in a common reference frame; The image is divided into P spatial regions (34) to generate a set of P image blocks (36) corresponding to the P spatial regions; The set of training images (38) in the common reference frame is combined by selecting an image patch from each of the P image patch sets for each training image in the common reference frame and combining the selected image patch into the training image in the common reference frame; Processing the training images in the public reference frame to generate the training set of anonymous images includes applying a statistical inverse space mapping to the training images in the public reference frame, wherein the statistical inverse space mapping is derived from a spatial mapping (33) from the images of the plurality of people to the reference image.

15. A non-transient computer-readable medium (26) storing instructions that can be executed by at least one electronic processor (20) to perform the method according to claim 14.

Images