Method for enriching a database of medical images intended for a clinical study

Abstract

The invention relates to a method for enriching a database of medical images for a clinical study, the method comprising the steps of: - retrieving (100) a medical image to be analyzed; - identifying (103) an organ type in the medical image to be analyzed by implementing an artificial intelligence model generated by training on a database of reference medical images for which the values of the predefined parameters are fixed; - selecting the predefined parameters of the image to be analyzed according to the identified organ type; - identifying (106) the predefined parameter values of the medical image to be analyzed by implementing the artificial intelligence model; - storing (108) the identified parameter values in metadata associated with the image to be analyzed.

Description

Description Title of the invention: Method for enriching a database of medical images intended for a clinical study

[0001] [The invention relates to the field of clinical studies using medical images, and in particular to systems for assisting in the consultation of such medical images in the context of the preparation of clinical studies.

[0002] Regulations governing the distribution of medical devices require prior marketing authorizations. Such authorizations are granted only after clinical studies have been conducted to determine the suitability of a medical device for use and under what conditions.

[0003] In particular, medical prostheses and their implantation methods must undergo clinical studies before marketing authorization is granted. These clinical studies rely, notably, on extensive databases of images illustrating experimental implantations of such medical prostheses. These images include, for example, X-rays of prosthesis implantations in various parts of the body, such as the foot, hip, shoulder, leg, or knee, taken from different angles.

[0004] The validity of a clinical study relies, in particular, on the analysis of a number of clinical cases. To facilitate this analysis, the images are associated with metadata. This metadata allows a researcher to sort and select clinical cases to effectively conduct the study.

[0005] Metadata entry is carried out by various practitioners or users. This metadata may be entered inconsistently, incompletely, or incorrectly. This leads to metadata imperfections that make using the clinical case database difficult. Furthermore, computing resources are limited in some environments, with terminals that may be nearing obsolescence. R012469 PCT Deposit Text

[0006] Document W02020020770 describes a method for managing a medical imaging system. This system adapts imaging parameters in real time based on artificial intelligence image analysis, in order to improve the image displayed to the user. To achieve this, the document retrieves a medical image for real-time analysis, identifies the muscle area within the image using artificial intelligence, and then applies corrections to the displayed image or to the imaging parameters.

[0007] Document US2022198784 describes the enrichment of a training image database for artificial intelligence. Using existing images, catheter models are, for example, inserted into modified copies of the images.

[0008] Document WO2023172621 describes a method for introducing envisaged prosthetic shapes into existing photographs.

[0009] The invention aims to resolve one or more of these drawbacks. The invention thus relates to a method for enriching a database of medical images intended for a clinical study, as defined in attached claim 1.

[0010] The invention also relates to variants of dependent claims. Those skilled in the art will understand that each of the features of variants of the description or of a dependent claim can be combined independently with the features of an independent claim, without thereby constituting an intermediate generalization.

[0011]

[0012] The invention also relates to a digital system comprising a digital processing device configured to implement a process as defined above.

[0013] Other features and advantages of the invention will become clear from the description given below, which is by way of example and not limitation, with reference to the accompanying drawings, in which: R012469 PCT Deposit Text

[0014] [Fig.1] is a schematic representation of a system for implementing a method for enriching a database of medical images according to the invention;

[0015] [Fig.2] is an example of a flowchart of a database enrichment process applied to images of implanted knee prostheses;

[0016] [Fig.3] is an example of a flowchart of a database enrichment process applied to images of implanted hip prostheses;

[0017] [Fig.4] is an example of a flowchart of a database enrichment process applied to images of implanted leg prostheses.

[0018] Figure 1 is a schematic representation of a system 1 for implementing a method for enriching a database of medical images according to the invention. System 1 comprises digital processing means known per se and an application system for implementing the method.

[0019] System 1 is specifically configured to implement various artificial intelligence models. For example, the AI ​​model was generated by training on a database 2 of reference medical images. The AI ​​model can be based on a convolutional neural network such as ResNet50, ResNet18, ResNet152, or VGG. For each of these reference medical images, the training database 2 contains metadata with values ​​assigned to various predefined parameters. These predefined parameters are identical for all images within the same category. Based on its AI model, System 1 is therefore able to identify values ​​for the predefined parameters of a medical image to be analyzed.

[0020] A database 3 of images to be analyzed is accessible to system 1. The images in database 3 may have missing metadata, incomplete metadata, erroneous metadata, or inconsistent metadata between images. A terminal 4 allows a user to read from database 3 or to enroll new images. A terminal 5 can provide a R012469 PCT Deposit Text A user-friendly human-machine interface is used to monitor the identification of predefined parameter values ​​and to monitor the storage of these values. Terminal 6 allows a user to consult database 3, based on metadata generated as defined below.

[0021] For the implementation of the enrichment process, system 1 retrieves a medical image to be analyzed from database 3. The image to be analyzed can also be submitted in real time to system 1.

[0022] By implementing its artificial intelligence models for image analysis, system 1 identifies values ​​of predefined parameters of the medical image to be analyzed, for example during steps 102 to 107 or 109, 110, or even 112 to 115. These predefined parameters are imposed by the artificial intelligence models.

[0023] Once the values ​​of the predefined parameters are identified, these values ​​are stored in metadata associated with the analyzed image. This storage can be performed in database 3, or in a database independent of the initial database 3.

[0024] The method according to the invention thus enables the automatic identification of values ​​for a number of predefined parameters. The resulting database provides metadata with consistent data for all images. A user can therefore benefit from a greater capacity for selecting clinical cases, with inherent reliability. Errors in assigning medical images can be overcome.

[0025] The predefined parameters will typically include the presence or absence of a prosthesis, the type of view of the image to be analyzed, and the part of the body corresponding to the image.

[0026] The process is particularly useful for analyzing radiographs of medical prostheses, especially implantations of knee, leg, hip, shoulder or foot prostheses. R012469 PCT Deposit Text

[0027] The flowchart in Figure 2 allows us to verify the process of an example of a database enrichment process 3 applied to images of implanted knee prostheses.

[0028] In step 100, a medical image is retrieved for analysis, and its metadata needs to be enriched. Image retrieval can be implemented, for example, by accessing database 3, such as a hospital's medical database. Image retrieval may occur after a user-defined filtering step among the images in database 3. After image retrieval, a preliminary normalization step (step 101) can be implemented. Normalization aims to ensure consistency across all images to be analyzed, making the artificial intelligence analysis more reliable and consistent across different images. Normalization step 101 may include converting the image to a predefined file format, converting the file to a predefined color scale (grayscale, RGB encoding, etc.).), resizing the image to a predefined dimension or cropping the image. A number of predefined parameter values ​​are then identified. These predefined parameters can, of course, be adapted depending on the type of organ being analyzed.

[0029] In step 102, a verification step is performed to determine if the image is indeed a medical image. The value returned and then stored for this predefined parameter could, for example, be 0 if the image is not medical in nature or 1 if the image is indeed medical in nature.

[0030] In step 103, the organ identified in the image is identified. The returned and stored value could be, for example, 'knee', 'hip', 'leg', 'foot', or 'elbow', depending on the organ identified in the image. Subsequent steps will depend on the type of organ identified. This control allows for the selection of a suitable decision tree, that is, an algorithm tailored to the organ identified in the image. The image parameters to be analyzed by the artificial intelligence model can then be defined based on the organ identified in step 103. R012469 PCT Deposit Text

[0031] Because of the selection of the appropriate algorithm, the process is particularly resource-efficient and allows the use of devices with limited capabilities, including personal computers that are not of the latest generation.

[0032] In step 104, it is determined whether the image is an assembly of several different images (sometimes images from different viewpoints are combined). The identified value is then stored.

[0033] In step 105, the type of shot taken is identified. The returned and stored value can, for example, be selected from a front view, a top view, or a side view.

[0034] In this case, subsequent processing will depend on the type of view identified in step 105. Here too, the algorithm implemented will depend on the type of view identified in step 105, thus saving the necessary processing resources.

[0035] Thus, to identify a top view, steps 106 to 108 will be implemented. In step 106, the type of prosthesis is identified by its laterality: is there a prosthesis on the left knee, the right knee, on both knees, or on neither knee? The corresponding value is then stored. In step 107, markers can be placed based on the identified prosthesis type. Areas of interest on the image can also be defined. In step 108, the metadata values ​​associated with the image are sent to database 3.

[0036] Thus, to identify a side view, steps 109 to 111 will be implemented. In step 109, the type of prosthesis is identified by its category: is it a total or partial prosthesis? The corresponding value is then stored. In step 110, markers can be placed based on the identified prosthesis type. Areas of interest on the image can also be defined. In step 111, the metadata values ​​associated with the image are sent to database 3. R012469 PCT Deposit Text

[0037] Thus, for the identification of a frontal view, steps 112 to 116 will be implemented. In step 112, the type of prosthesis is identified by its category: is it a total or partial prosthesis? The corresponding value is then stored. In step 113, the number of organs present on the image is noted. If two knees are identified on the image, this can be considered an anomaly for this type of view. In step 114, the type of prosthesis is identified by its laterality: is it a left or right prosthesis? In step 115, markers can be placed based on the identified prosthesis type. Areas of interest on the image can also be defined. In step 116, the metadata values ​​associated with the image are sent to database 3.

[0038] By providing a human-machine interface during the analysis process, the values ​​identified by the artificial intelligence processing can be presented to the user. These values ​​can be submitted to the user as suggestions. The user can then choose to modify these values ​​at their discretion, in order to save values ​​other than those suggested.

[0039] Furthermore, when returning the images to the user, duplicates can be advantageously removed.

[0040] The flowchart in Figure 3 allows us to verify the process of an example of a database enrichment process applied to images of implanted hip prostheses.

[0041] In step 200, a medical image is retrieved for analysis, and its metadata must be enriched. After image retrieval, a preliminary normalization step (step 201) can be implemented, similar to step 101. A number of predefined parameter values ​​are then identified. These predefined parameters can, of course, be adapted according to the type of organ being analyzed.

[0042] In step 202, a verification step is performed to confirm that the image is indeed a medical image. The value returned and then stored for this predefined parameter R012469 PCT Deposit Text For example, it could be worth 0 if the image is not of a medical nature or 1 if the image is indeed of a medical nature.

[0043] In step 203, the organ shown in the image is identified. The value returned here will be 'hip', to illustrate the corresponding algorithm. Subsequent steps will depend on the type of organ identified. This control allows us to select a suitable decision tree, that is, an algorithm adapted to the organ identified in the image. The image parameters to be analyzed by the artificial intelligence models can then be defined according to the organ identified in step 203, which in this case is specific to a hip.

[0044] In step 204, it is identified whether the image is an assembly of several disparate images. The identified value is then stored.

[0045] In step 205, the type of shot taken is identified. The value returned and stored can, for example, be selected from a front view with two hips, a front view with one hip, or another view.

[0046] In this case, subsequent processing will depend on the type of view identified in step 205. Here too, the algorithm implemented will depend on the type of view identified in step 205, thus saving the necessary processing resources.

[0047] Thus, to identify a frontal view with two hips, steps 206 to 208 will be implemented. In step 206, the prostheses are identified. The returned values ​​can be, for example, 'no', 'left', 'right', or 'left and right'. In step 207, markers can be placed based on the type of prosthesis identified. Areas of interest on the image can also be defined. In step 208, the metadata values ​​associated with the image are sent to database 3.

[0048] Thus, to identify a frontal view with a hip, steps 209 to 211 will be implemented. In step 209, the presence of a prosthesis is checked. The returned values ​​can be, for example, 'yes' or 'no'. In step 210, the laterality of the hip is identified. The returned values ​​can be 'left' or 'right'. Areas of interest in the image may also be R012469 PCT Deposit Text defined. During step 211, the metadata values ​​associated with the snapshot are returned to database 3.

[0049] Thus, for the identification of another view, for example oblique, false profile or lateral, other steps may be implemented and will not be detailed here.

[0050] The flowchart in Figure 4 allows us to verify the process of an example of a database enrichment process 3 applied to leg radiographs.

[0051] In step 300, a medical image is retrieved for analysis, and its metadata must be enriched. After image retrieval, a preliminary normalization step 301, similar to step 101, can be implemented. A number of predefined parameter values ​​are then identified. These predefined parameters can, of course, be adapted according to the type of organ being analyzed.

[0052] In step 302, a verification step is performed to determine if the image is indeed a medical image. The value returned and then stored for this predefined parameter could, for example, be 0 if the image is not medical in nature or 1 if the image is indeed medical in nature.

[0053] In step 303, an identification step is performed to determine the organ shown in the image. The value returned here will be 'leg', to illustrate the corresponding algorithm. The subsequent steps will therefore depend on the type of organ identified. This control allows us to select a suitable decision tree, that is, an algorithm adapted to the organ identified in the image. The image parameters to be analyzed based on the artificial intelligence model can then be defined according to the organ identified in step 303, in this case, specific to a leg.

[0054] In step 304, it is determined whether the image is an assembly of several disparate images. The identified value is then stored.

[0055] In step 305, the type of image capture is identified. The returned and stored value can, for example, be selected from a view of R012469 PCT Deposit Text A front view with two legs, a front view with one left leg, a front view with one right leg, or a side view.

[0056] In this case, subsequent processing will depend on the type of view identified in step 305. Here too, the algorithm implemented will depend on the type of view identified in step 305, thus saving the necessary processing resources.

[0057] Thus, subsequent steps may consist of identifying the presence of a prosthesis, identifying the organ equipped with a prosthesis, or identifying the laterality of the prosthesis. R012469 PCT Deposit Text

Claims

Demands

1. [Method for enriching a database of medical images intended for a clinical study, characterized in that it comprises the steps of: -retrieval (100) of a medical image to be analyzed, of an implantation of a surgical prosthesis of the knee, hip or leg; -identification (103) of a type of organ in the medical image to be analyzed by implementing an artificial intelligence model that has been generated by training on a database of reference medical images whose predefined parameter values ​​are fixed; -selection of predefined parameters, depending on the type of organ identified, these predefined parameters to be analyzed in the image to be analyzed, the predefined parameters including the presence or absence of a prosthesis and the type of view of the image to be analyzed; -identification (106) of the values ​​of the selected predefined parameters, in the medical image to be analyzed by implementing the artificial intelligence model, -storage (108) of the identified values ​​of the parameters in metadata associated with the image to be analyzed.

2. An enrichment method according to claim 1, wherein the predefined parameters further include a parameter selected from the group including the number of organs present on the image, parameters specific to the type of view of the image, or the type of prosthesis present on the image.

3. Enrichment method according to claim 1 or 2, wherein the method includes a preliminary step of normalizing the image to be analyzed, comprising any one of the following steps: -conversion of the image into a color scale; -Resizing the image to a predefined dimension.

4. An enrichment method according to any one of the preceding claims, comprising a step of querying a user via a human-machine interface, including submitting to the user a proposed value for a parameter R012469 PCT Deposit Text predefined and waiting for confirmation of the value by the user, prior to memorizing the identified value of the parameter.

5. An enrichment method according to any one of the preceding claims, comprising repeating the steps of claim 1 for a succession of medical images to be analyzed, so as to constitute a database of images associated with the stored metadata, this database being accessible through queries based on the values ​​of predefined parameters.

6. Digital system comprising a digital processing device configured to implement a method according to any one of the preceding claims. R012469 PCT Deposit Text

Images