Method and device for analyzing x-ray image for musculoskeletal disease

Abstract

A device according to an embodiment of the present disclosure comprises: at least one memory storing instructions; and at least one processor. The at least one processor executes the instructions to perform the processes of: generating a plurality of pyramid feature maps in a multilayer form by applying a pyramid model to an input X-ray image; processing the plurality of generated pyramid feature maps by fusing a binary embedding code thereto; and generating a summed feature by summing the plurality of processed pyramid feature maps according to a predetermined data set, and classifying the summed feature.

Description

Method and device for analyzing X-ray images for musculoskeletal diseases

[0001] The present disclosure relates to a method and device for analyzing X-ray images for musculoskeletal diseases.

[0002] The material described in this section merely provides background information for the present invention and does not constitute prior art.

[0003] The musculoskeletal system, comprising muscles and the bones that form the framework, comprises approximately 40% of the body and functions to maintain movement, support, and balance. Musculoskeletal disorders are a general term for health problems affecting muscles, tendons, nerves, blood vessels, and surrounding body tissues. Examples include carpal tunnel syndrome, tendonitis, thoracic outlet syndrome, and tension neck syndrome.

[0004] Musculoskeletal disorders can occur in young and middle-aged individuals, the mainstays of economic activity, due to repetitive activities and prolonged awkward postures. However, most are degenerative diseases associated with aging. Depending on the condition, these musculoskeletal disorders often present with pain during work or at rest.

[0005] Because such musculoskeletal disorders are the most common cause of severe long-term pain and disability not only in young and middle-aged adults but also in the elderly over 65 years of age, there is a growing demand for methods and devices that accurately and quickly analyze X-ray images for the diagnosis of musculoskeletal disorders.

[0006] The present disclosure provides a method and device for accurately and quickly analyzing X-ray images for diagnosing musculoskeletal diseases.

[0007] The present disclosure provides a method and device for improving recognition performance by applying an image related to a musculoskeletal disease to a pyramid model.

[0008] The present disclosure provides a method and device for improving recognition performance by binary encoding images of different sizes.

[0009] The problems to be solved by the present invention are not limited to the problems mentioned above, and other problems not mentioned will be clearly understood by those skilled in the art from the description below.

[0010] According to one aspect of the present disclosure, there is provided a device for analyzing an X-ray image for a musculoskeletal disease, comprising: at least one memory for storing commands; and at least one processor, wherein the at least one processor is configured to perform a process of applying a pyramid model to an input X-ray image to generate a plurality of pyramid feature maps in a multi-layered form by executing the commands; a process of fusing binary embedding codes to the generated plurality of pyramid feature maps and processing them; and a process of combining the processed plurality of pyramid feature maps according to a predetermined data set to generate a combined feature, and classifying the combined feature.

[0011] According to another aspect of the present disclosure, a computer-implemented method is provided, comprising: a step of applying a pyramid model to an input X-ray image to generate a plurality of pyramid feature maps in a multi-layered form; a step of processing the generated plurality of pyramid feature maps by fusing a binary embedding code; and a step of generating a combined feature by combining the processed plurality of pyramid feature maps according to a predetermined data set; and a step of classifying the combined feature.

[0012] The present disclosure can improve recognition performance by applying images of different sizes to a pyramid model.

[0013] The present disclosure can improve recognition performance by binary encoding images of different sizes.

[0014] The effects of the present disclosure are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

[0015] FIG. 1 is a block diagram of a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure.

[0016] FIG. 2 is a flowchart illustrating the operation of a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure.

[0017] FIG. 3 is a flowchart illustrating the operation of a binary encoding block in a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure.

[0018] FIG. 4 and FIG. 5 are exemplary diagrams showing the recognition performance of a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure.

[0019] FIG. 6 and FIG. 7 are other exemplary diagrams showing the recognition performance of a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure.

[0020] FIG. 8 and FIG. 9 are further exemplary diagrams showing the recognition performance of a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure.

[0021] Hereinafter, some embodiments of the present disclosure will be described in detail with reference to exemplary drawings. When designating components in each drawing, it should be noted that, where possible, identical components are given the same reference numerals, even if they appear in different drawings. Furthermore, when describing the present disclosure, detailed descriptions of related known structures or functions will be omitted if they are deemed to obscure the gist of the present disclosure.

[0022] In addition, terms such as first, second, A, B, (a), (b), etc. may be used to describe components of the present disclosure. These terms are only intended to distinguish the components from other components, and the nature, order, or sequence of the components are not limited by the terms. Throughout the specification, when a part is said to "include" or "have" a component, this does not mean that other components are excluded, but rather that other components may be further included, unless specifically stated otherwise. In addition, terms such as "part" and "module" described in the specification mean a unit that processes at least one function or operation, and this may be implemented by hardware, software, or a combination of hardware and software.

[0023] The detailed description set forth below, together with the accompanying drawings, is intended to explain exemplary embodiments of the present disclosure and is not intended to represent the only embodiments in which the present disclosure may be practiced.

[0024] The recognition performance in this specification may refer to the recognition performance of musculoskeletal diseases or other various forms of recognition performance, but is not limited to a specific form.

[0025] The images in this specification may include X-ray images or other various images, and are not limited to a specific form.

[0026] FIG. 1 is a block diagram of a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure.

[0027] A device for analyzing an X-ray image for a musculoskeletal disease according to an embodiment of the present disclosure may be configured to include a pyramid block (100), a binary encoding block (110), a combining block (120), a classification block (130), etc., as shown in FIG. 1.

[0028] The output result in the pyramid block (100) can be called a “pyramid feature map”, “feature”, or “feature map”.

[0029] A pyramid block (100) can extract multiple layer-specific features or multiple pyramid feature maps (the first feature, the second feature, the third feature, and the fourth feature of FIG. 1) in a multi-layered form by applying a pyramid model (the pyramid model can be referred to as a pyramid layer, a feature pyramid layer, or a feature pyramid model) (101, 102, 103, 104) to an X-ray image. The pyramid model can include a model that can train and output the X-ray image by n layers, as shown in FIG. 1. The multiple pyramid feature maps can have different sizes (or scales).

[0030] The pyramid model is a feature pyramid, and the FOV (Field of View) that the network sees is different for each pixel of the feature map for each layer depending on the size of the affected area, and it can include a network model that uses different optimal resolution features combined.

[0031] The binary encoding block (110) can process a plurality of pyramid feature maps by flattening and then binary encoding them. That is, the binary encoding block (110) can be configured to fuse a dataset binary embedding having a length of 2 into a plurality of pyramid feature maps. Combining a dataset binary embedding having a length of 2 into a plurality of pyramid feature maps means extracting a plurality of pyramid feature maps that are processed (or vectorized) by combining binary codes (e.g., 00, 01, 10, 11) into each of the plurality of pyramid feature maps.

[0032] The binary encoding block (110) can make the lengths of the plurality of flattened one-dimensional matrices the same when the lengths of the plurality of flattened one-dimensional matrices are different for each layer.

[0033] The binary encoding block (110) can easily distinguish between classes by combining a dataset binary embedding of length 2 into multiple pyramid feature maps, thereby widening the distribution shift margin of the multiple pyramid feature maps.

[0034] An embodiment of the present disclosure can increase the discriminancy between data by placing an embedding binary encoder within a binary encoding block (110) and placing a feature that can expand the distribution shift between each dataset by the number of data digits, thereby widening the data distance in the feature space between classes. The embedding binary encoder can facilitate dataset distinction, image type identification, or identification of multiple generated pyramid feature maps. It can also be used in a simple binary embedding manner depending on the system-specific required size of the dataset or feature dimension.

[0035] After combining, the combined block (120) can be combined immediately since the layer of the pyramid feature map is 15253504+2 for Stage 1 and 3.

[0036] On the other hand, after combining, the combining block (120) cannot be combined directly because the layer of the pyramid feature map is 7626752+2 in the case of Stage 2 and 4. Therefore, in the case of Stage 2 and 4, the combining block (120) can be combined after passing through the FC (Fully Connected) layer once more (at this time, weights are applied).

[0037] The combining block (120) can combine (sum) the outputs of the binary encoding block (110) and output them as a single image. That is, the combining block (120) can combine four pyramid feature maps of length 2 according to a predetermined dataset and output a combined feature. The predetermined dataset can be determined according to the size and shape of the affected area. The predetermined dataset is, for example, {4 th{Feature}, {4 th + 2 nd {Feature}, {4 th + 3 rd + 2 nd {Feature}, {4 th +3 rd + 2 nd + 1 st It can be expressed as {characteristic}.

[0038] The combining block (120) according to the embodiment of the present disclosure may apply "sum" rather than "concatenate" when combining. The reason for this is as follows.

[0039] An embodiment of the present disclosure can determine a single conclusion from a single image as normal or abnormal as the output of the classification block (130). In this case, in the case of a task of detecting or segmenting object boxes of various sizes according to the image scale by placing features of various FoV sizes, a task that requires recognition of multiple instances to succeed, unlike the operation of the classification block, can be included.

[0040] On the other hand, the classification block (130) only needs to output normal or abnormal regardless of the size of the affected area. Therefore, only one loss function can be set to optimize the model's output. However, there are multiple ways to arrive at this conclusion in the final layer. Therefore, combining multiple features by summing rather than concatenating them into a single output can yield better performance.

[0041] The classification block (130) can output class-specific probability values ​​for the combined features of the combining block (120).

[0042] More specifically, the classification block (130) may be configured to pass the combined feature of the combination block (120) through a binary cross entropy loss layer to secure classification performance. Afterwards, the classification block (130) finally obtains probability values ​​of two classes when the combined feature passes through a logit layer, and a normal or abnormal decision can be made for a musculoskeletal disease based on the probability values.

[0043] Therefore, the output of the combination block (120) can be used to determine the presence or absence of a musculoskeletal disease in the classification block (130).

[0044] The operation of the pyramid block (100), binary encoding block (110), combination block (120), and classification block (130) is described in more detail as follows.

[0045] As shown on the left side of Fig. 1, the embodiment of the present disclosure uses the TResNet baseline, and after the output feature dimension (or pyramid feature map) of each feature pyramid layer (101, 102, 103, 104) passes through the flattening section, the size of the feature map, the number of channels, and the feature dimension in each of Stages 1 to 4 can be expressed as shown in Table 1.

[0046]

[0047] When a pyramid feature map is generated, the binary encoding block (110) combines a dataset binary embedding of length 2 into each of a plurality of pyramid feature maps to widen the distribution shift margin of the feature, thereby facilitating distinction between classes and image identification.

[0048] The joining unit (121) can combine four pyramid feature maps of length 2.

[0049] After combining, the combined block (120) can be combined immediately since the layer of the pyramid feature map is 15253504+2 for Stage 1 and 3.

[0050] On the other hand, after combining, the combined block (120) cannot be combined immediately because the layer of the pyramid feature map is 7626752+2 in the case of Stage 2 and 4. Therefore, in the case of Stage 2 and 4, the combined block (120) can be combined after passing through the FC (Fully Connected) layer once more (at this time, weights are applied).

[0051] The classification block (130) can cause the output value of the above-mentioned combination unit (121) to pass through a binary cross entropy loss layer, and finally, can cause it to pass through a logit layer to output probability values ​​of two classes, and can use the probability values ​​to determine whether a musculoskeletal disease is normal or abnormal.

[0052] In an embodiment of the present disclosure, a head classifier optimized for pyramid features is designed to improve recognition performance. Specifically, a method of setting the dimension of features from block layers for each feature resolution to the same 2 by probability processing them as binary logit values ​​and then averaging them can improve performance compared to a method of concatenating them to 8 and then adding an FC layer to make them 2.

[0053] In the embodiments of the present disclosure, three methods are proposed to improve recognition performance.

[0054] First, the device according to the embodiment of the present disclosure may include a feature map (pyramid feature map) in the form of a multi-layer (n-layer) pyramid that takes into account various influence area sizes.

[0055] Second, the device according to the embodiment of the present disclosure can apply a dataset binary encoding process that performs binary embedding and feature fusion for each data type.

[0056] The third feature merging step creates the model output by adding or connecting binary features for each layer.

[0057] Since only one classification result needs to be derived from a single image, the total loss can be structured as a single loss system. In quantitative analysis, we compared the performance of each of the four measurement features for each feature pyramid combination. Four combinations can be considered.

[0058] 4 th , 4 th + 2 nd , 4 th + 3 rd + 2 nd , 4 th +3 rd + 2 nd + 1 st , a total of four types of multi-layer (n-layer) features can be considered, as shown in Table 2. Intermediate feature fusion can apply a binary embedding method that can be used as a dataset discriminator.

[0059] In the present disclosure, the MURA dataset was used, with only one binary code, '10', added for intermediate feature and embedding fusion. Tables 3 and 4 demonstrate high performance in terms of CK Score and Accuracy.

[0060] The evaluation metrics used in Table 2 may include CK (Cohen-Kappa) score and Accuracy, and the evaluation metrics used in Tables 3 to 6 may include CK score, Accuracy, AUPRC (Aera Under the Precision-Recall Curve), and AUC (Area Under the Curve).

[0061] The above evaluation index may be an index that measures the normal or abnormal classification performance for musculoskeletal diseases during classification in the classification block (130).

[0062]

[0063] Table 3 shows the case when use_binary_enc=0, use_concate=0 was used (when binary codes 0, 0 were used), Table 4 shows the case when use_binary_enc=1, use_concate=0 was used (when binary codes 1, 0 were used), Table 5 shows the case when use_binary_enc=1, use_concate=1 was used (when binary codes 1, 1 were used), and Table 6 shows the case when use_binary_enc=0, use_concate=1 (when binary codes 0, 1 were used). It can be seen that the performance is highest when the four features are combined in Tables 4 and 5 than in Tables 5 and 6.

[0064] According to Tables 3 and 4, when all four features were used, the CK score and accuracy values ​​were higher than when the other three features were combined. However, for AUPRC and AUC, the performance of the remaining three feature combinations was almost similar to that of the TResNet baseline alone.

[0065] In summary, we can see that when four features are commonly used across all evaluation metrics, performance is improved compared to when only the TResNet baseline is used.

[0066]

[0067]

[0068] The feature combination set of the Feature Pyramid model optimized for musculoskeletal data is {(1, 2, 3, 4) or (2, 4) + binary encoding block + combination block}, as shown in Tables 3 and 4 above. Since this is the result for musculoskeletal data, for other diseases, different combinations of features may be optimal depending on the size and shape of the affected area related to the disease, so it goes without saying that the optimal combination can be found through additional experiments.

[0069]

[0070]

[0071] FIG. 2 is a flowchart illustrating the operation of a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure.

[0072] A device for analyzing X-ray images for musculoskeletal diseases can receive X-ray images at step 210.

[0073] A device for analyzing X-ray images for musculoskeletal diseases can extract multiple pyramid feature maps (the first feature, the second feature, the third feature, and the fourth feature in Fig. 1) in a multi-layered form by applying a pyramid model to the X-ray image in step 220.

[0074] The device for analyzing X-ray images for musculoskeletal diseases can flatten and binary-encode the plurality of pyramid feature maps, respectively, in step 230. The plurality of pyramid feature maps can be converted into a one-dimensional matrix by flattening. The binary encoding can be configured to combine a dataset binary embedding (binary embedding code) with the plurality of pyramid feature maps.

[0075] A device for analyzing X-ray images for musculoskeletal disorders can output binary encoded outputs by summing them in 240 steps.

[0076] A device for analyzing X-ray images for musculoskeletal diseases can output classified combined outputs in 250 steps.

[0077] FIG. 3 is a flowchart illustrating the operation of a binary encoding block in a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure.

[0078] The binary encoding block (110) can flatten and output multiple pyramid feature maps per layer in step 310.

[0079] The binary encoding block (110) can output a binary embedding (binary code) by combining multiple pyramid feature maps flattened in 320 steps.

[0080] The binary encoding block (110) can output a plurality of combined pyramid feature maps of length 2 in 330 steps.

[0081] The features described in FIGS. 4 to 9 refer to pyramid feature maps.

[0082] FIGS. 4 and 5 are exemplary diagrams illustrating the recognition performance of a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure. FIGS. 4 and 5 are displayed as forearm heat maps using Grad-CAM (Gradient-weighted Class Activation Mapping) to verify the recognition performance.

[0083] Figure 4 (a) shows the original image.

[0084] Figure 4 (b) is 4 th Indicates a case where a feature is used.

[0085] Figure 5 (a) is 2 nd+ 4 th Indicates a case where a feature is used.

[0086] 2 nd + 4 th As shown, it can be seen that the recognition performance was superior when two features were used.

[0087] Figure 5 (b) is 1 st + 2 nd + 3 rd + 4 th Indicates a case where a feature is used. 1 st + 2 nd + 3 rd + 4 th As shown in Fig. 4(a) (original image) and Fig. 4(b) (4) using four features. th It can be seen that the recognition performance is superior to that of using a single feature, as in the case of the present disclosure. Here, the recognition performance according to the embodiment of the present disclosure can be achieved on the MURA (musculoskeletal radiography) X-ray image dataset by grafting the pyramid feature concept based on TResNet, which shows high performance in ImageNet classification. Here, MURA can be used as an imaging method for viewing musculoskeletal abnormalities.

[0088] It can be seen that Fig. 5 (a) and Fig. 5 (b) activate a wider range of fracture sites overall by using more features compared to Fig. 4 (b).

[0089] Additionally, while Figure 5(a) activates two areas, Figure 5(b) activates a single area more broadly, which may help yield more accurate clinical diagnosis. Furthermore, additional data may help yield more accurate clinical diagnoses of the humerus and wrist.

[0090] FIGS. 6 and 7 are further exemplary diagrams illustrating the recognition performance of a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure. FIGS. 6 and 7 are displayed as heatmaps of the humerus using Grad-CAM to verify the recognition performance.

[0091] Figure 6 (a) shows the original image.

[0092] Figure 6 (b) is 4 th Indicates a case where a feature is used.

[0093] Figure 7 (a) is 2 nd + 4 th Indicates a case where a feature is used.

[0094] 2 nd + 4 th As shown, it can be seen that the recognition performance was superior when two features were used.

[0095] Figure 7 (b) is 1 st + 2 nd + 3 rd + 4 th Indicates a case where a feature is used. 1 st + 2 nd + 3 rd + 4 th As shown in Fig. 6 (a) (original image) and Fig. 6 (b) (4) using four features. th It can be seen that the recognition performance was superior to that of using a single feature (such as ).

[0096] It can be confirmed that Fig. 7 (a) and Fig. 7 (b) have activated the fracture site so that the fracture site can be recognized more clearly, although in a narrower range compared to Fig. 6 (b).

[0097] Additionally, while (a) of Fig. 7 activates two locations, (b) of Fig. 7 activates one location intensively, which can help to provide more accurate findings in clinical diagnosis.

[0098] FIGS. 8 and 9 are further exemplary diagrams illustrating the recognition performance of a device for analyzing X-ray images for musculoskeletal diseases according to an embodiment of the present disclosure. FIGS. 8 and 9 are represented as wrist heatmaps using Grad-CAM to verify the recognition performance.

[0099] Figure 8 (a) shows the original image.

[0100] Figure 8 (b) is 4 th Indicates a case where a feature is used.

[0101] Figure 9 (a) is 2 nd + 4 th Indicates a case where features are used. 2 nd + 4 th As shown, it can be seen that the recognition performance was superior when two features were used.

[0102] Figure 9 (b) is 1 st + 2 nd + 3 rd + 4 th This shows the case where it is applied. 1 st + 2 nd + 3 rd + 4 th As in the case of using four features, Fig. 8 (a) (original image) and Fig. 8 (b) (4 th It can be seen that the recognition performance was superior to that of using a single feature (such as ).

[0103] It can be confirmed that Fig. 9 (a) and Fig. 9 (b) have activated the fracture site so that a more distinct fracture site can be recognized, although the range is narrower than Fig. 8 (b).

[0104] Additionally, while (a) of Fig. 9 activates one location, (b) of Fig. 9 activates two locations intensively, which can help to provide more accurate findings in clinical diagnosis.

[0105] Various implementations of the systems and techniques described herein may be implemented as digital electronic circuits, integrated circuits, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations of one or more computer programs executable on a programmable system. The programmable system includes at least one programmable processor (which may be a special purpose processor or a general purpose processor) coupled to receive data and instructions from and transmit data and instructions to a storage system, at least one input device, and at least one output device. Computer programs (also known as programs, software, software applications, or code) include instructions for the programmable processor and are stored on a "computer-readable recording medium."

[0106] A computer-readable recording medium includes any type of recording device that stores data that can be read by a computer system. Such a computer-readable recording medium may be a non-volatile or non-transitory medium such as a ROM, CD-ROM, magnetic tape, floppy disk, memory card, hard disk, magneto-optical disk, storage device, and may further include a transitory medium such as a data transmission medium. Furthermore, the computer-readable recording medium may be distributed across network-connected computer systems, so that computer-readable code can be stored and executed in a distributed manner.

[0107] Various implementations of the systems and techniques described herein can be implemented by a programmable computer. The computer includes a programmable processor, a data storage system (including volatile memory, non-volatile memory, or other types of storage systems, or a combination thereof), and at least one communication interface. For example, the programmable computer can be one of a server, a network appliance, a set-top box, an embedded device, a computer expansion module, a personal computer, a laptop, a personal data assistant (PDA), a cloud computing system, or a mobile device.

[0108] The above description is merely an example of the technical idea of ​​the present embodiment, and those skilled in the art will appreciate that various modifications and variations can be made without departing from the essential characteristics of the present embodiment. Therefore, the present embodiments are not intended to limit the technical idea of ​​the present embodiment, but rather to explain it, and the scope of the technical idea of ​​the present embodiment is not limited by these embodiments. The scope of protection of the present embodiment should be interpreted by the claims below, and all technical ideas within a scope equivalent thereto should be interpreted as being included in the scope of rights of the present embodiment.

[0109] CROSS-REFERENCE TO RELATED APPLICATION

[0110] This patent application claims priority to Korean Patent Application No. 10-2023-0093879, filed July 19, 2023, which is incorporated herein by reference in its entirety.

Claims

1. A device for analyzing X-ray images for musculoskeletal diseases. At least one memory for storing instructions; and at least one processor, At least one processor executes the instructions, A process of generating multiple pyramid feature maps in a multi-layered form by applying a pyramid model to an input X-ray image; A process of processing the plurality of generated pyramid feature maps by combining (fusion) a binary embedding code; and A device configured to perform a process of combining (summing) the processed plurality of pyramid feature maps according to a predetermined data set to generate a combined feature and classifying the combined feature.

2. In the first paragraph, the pyramid model, A device that is a model that trains the input X-ray image in n layers.

3. In paragraph 1, A device wherein the above plurality of pyramid feature maps have different sizes.

4. In the first paragraph, the processing process is: A device comprising a process of converting the generated plurality of pyramid feature maps into a one-dimensional matrix by flattening them prior to combining the binary embedding codes.

5. In paragraph 4, The above processing process is, A device comprising a process of adjusting the lengths of the one-dimensional matrices to be the same when the lengths of the flattened plurality of one-dimensional matrices are different for each layer.

6. In paragraph 1, The above binary embedding code is used to identify the plurality of generated pyramid feature maps.

7. In the first paragraph, the device, wherein the predetermined data set is determined according to the size and shape of the refund.

8. In the first paragraph, the classification process is: A device comprising a process for determining normality or abnormality for the musculoskeletal disease based on class-specific probability values ​​for the above combined features.

9. As a computer implementation method, A process of generating multiple pyramid feature maps in a multi-layered form by applying a pyramid model to an input X-ray image; A process of processing the plurality of generated pyramid feature maps by combining (fusion) a binary embedding code; and A computer-implemented method comprising a process of combining (summing) the processed plurality of pyramid feature maps according to a predetermined data set to generate a combined feature and classifying the combined feature.

10. In paragraph 9, the pyramid model, A computer implementation method for training a model for n layers of the above input X-ray images.

11. In paragraph 9, A computer-implemented method wherein the plurality of pyramid feature maps have different sizes.

12. In paragraph 9, the processing process is: A computer-implemented method, comprising a process of flattening the generated plurality of pyramid feature maps and converting them into a one-dimensional matrix prior to combining the binary embedding codes.

13. In paragraph 13, The above processing process is, A computer-implemented method, comprising a process of making the lengths of the flattened plurality of one-dimensional matrices the same when the lengths of the one-dimensional matrices are different for each layer.

14. In paragraph 9, A computer-implemented method wherein the binary embedding code is used to identify the plurality of generated pyramid feature maps.

15. A computer-implemented method according to claim 9, wherein the predetermined data set is determined based on the size and shape of the affected area.

16. In paragraph 9, the classification process is: A computer-implemented method comprising a process of determining normality or abnormality for the musculoskeletal disease based on class-specific probability values ​​for the above-mentioned combined features.

17. A computer-readable recording medium having stored thereon a command, wherein the command, when executed by the computer, causes the computer to execute each process included in the computer implementation method according to claim 8.

Images