Deep Learning-Based Diabetic Retinopathy Lesion Segmentation Method

Abstract

The present invention discloses a method for diabetic retinopathy lesion segmentation based on deep learning, which relates to the field of image enhancement technology. A retinal image is determined, the retinal image is divided to obtain a plurality of image patches, the plurality of image patches are subjected to hierarchical processing to obtain a plurality of image slices; image enhancement is performed based on the image slices, and the image slices after image enhancement are fused to obtain a target retinal image; the target retinal image is processed based on deep learning technology to obtain a lesion segmentation result. A retinal image to be subjected to lesion segmentation is acquired, and the retinal image is divided into a plurality of image patches according to a preset size. By performing hierarchical processing on the retinal image for image enhancement, the present invention can perform local analysis and enhancement of the image in a more detailed and non-intrusive manner, ensure the accuracy after image enhancement, ensure the details of image enhancement, make the image contrast clearer, and facilitate the recognition and segmentation of subsequent convolutional networks.

Description

Technical Field

[0001] The present invention relates to the technical field of image enhancement, and particularly to a method for segmenting diabetic retinopathy lesions based on deep learning. Background Art

[0002] Diabetes is a common chronic disease. With the continuous increase in the number of global diabetes patients, the incidence of diabetic retinopathy (DR) is also rising. The traditional diagnosis of diabetic retinopathy lesions mainly relies on ophthalmologists to manually observe and judge through fundus images. However, this method has certain limitations. On the one hand, the interpretation of fundus images requires professional knowledge and rich experience. The number of ophthalmologists is relatively limited, and the training period is long, making it difficult to meet the diagnostic needs of a large number of patients. Therefore, the participation of deep learning technology can assist doctors in diagnosis. Although deep learning technology can provide certain help, the quality of the images also seriously affects the output results of the deep learning model. It is impossible to guarantee the quality of the images input into the deep learning model, which is not convenient for the lesion segmentation of the deep learning model. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for segmenting diabetic retinopathy lesions based on deep learning to solve the deficiencies in the background art.

[0004] To achieve the above purpose, the present invention provides the following technical solution: A method for segmenting diabetic retinopathy lesions based on deep learning, including the following steps:

[0005] Determine a retinal image, divide the retinal image into multiple image patches, and perform hierarchical processing on the multiple image patches to obtain multiple image slices;

[0006] Perform image enhancement based on the image slices, and fuse the image slices after image enhancement to obtain a target retinal image;

[0007] Process the target retinal image based on deep learning technology to obtain a lesion segmentation result.

[0008] In a preferred embodiment, the step of determining a retinal image, dividing the retinal image into multiple image patches, and performing hierarchical processing on the multiple image patches to obtain multiple image slices includes:

[0009] Obtain a retinal image that needs to be segmented for lesions, and divide the retinal image into multiple image patches according to a preset size;

[0010] Perform hierarchical processing on the multiple image patches according to brightness to obtain multiple image slices.

[0011] In a preferred embodiment, the step of performing hierarchical processing on multiple image blocks according to brightness to obtain multiple image slices includes:

[0012] Correspondingly configure multiple corresponding image beds for multiple image blocks, where an image bed includes multiple spatial planes;

[0013] Formulate a hierarchical brightness value range, which includes multiple preset hierarchical brightness ranges, and perform feature extraction on the image blocks according to the hierarchical brightness value range to obtain feature information corresponding to multiple preset hierarchical brightness ranges;

[0014] One-to-one place multiple pieces of feature information on the image bed respectively, and record the contours of the feature information on the image bed through the image bed to obtain multiple image slices.

[0015] In a preferred embodiment, the step of one-to-one placing multiple pieces of feature information on the image bed respectively and recording the contours of the feature information on the image bed through the image bed includes:

[0016] One-to-one place multiple pieces of feature information on the image bed in sequence according to the numerical sorting of the hierarchical brightness ranges where they are located, and the image beds are arranged in sequence;

[0017] Set the brightness interval range of the corresponding feature information on multiple spatial planes of the image bed, extract and layer the feature information within the brightness interval range and place it on multiple spatial planes, and mark the brightness of the feature information on the spatial planes. Among them, multiple photosensors are arranged on multiple spatial planes, and the brightness value and the corresponding range of the feature information are recorded through the photosensors.

[0018] In a preferred embodiment, the step of performing image enhancement based on the image slices and fusing the image slices after image enhancement to obtain a target retinal image includes:

[0019] Correspondingly formulate corresponding enhancement rules for multiple image blocks, where the enhancement rules include a brightness interval and a corresponding contrast enhancement difference value;

[0020] Regulate the photosensors in the image slices based on the enhancement rules to obtain a simulated image slice;

[0021] Use the simulated image slices that meet the preset conditions as simulated image slices, and assign the brightness and the corresponding range of the photosensors corresponding to the simulated image slices to the feature information on the spatial planes to obtain target image slices;

[0022] Overlap multiple target image slices to obtain a target image block, and splice multiple target image blocks corresponding to their positions in the retinal image to obtain a target retinal image.

[0023] In a preferred embodiment, the step of regulating the photosensors in the image patch based on the enhancement rules to obtain the image patch after simulation includes:

[0024] According to the brightness value and the corresponding range of the characteristic information recorded in the photosensor, obtain the brightness interval corresponding to the characteristic information and the existing contrast difference value as the adjustment base, and match the corresponding enhancement rules according to the adjustment base;

[0025] Perform brightness simulation regulation on the photosensors in the image patch according to the matched enhancement rules, and load the brightness regulated by the photosensors into the image patch to obtain the image patch after simulation.

[0026] In a preferred embodiment, the step of taking the image patch after simulation that meets the preset conditions as the simulated image patch, and assigning the brightness and the corresponding range of the photosensors corresponding to the simulated image patch to the characteristic information on the spatial plane to obtain the target image patch includes:

[0027] Formulate preset conditions, where the preset conditions are that the difference range of the contour coincidence between the image patch after simulation and the image patch satisfies F;

[0028] Take the image patch after simulation that meets the preset conditions as the simulated image patch, and adjust the brightness and the corresponding range of the photosensors corresponding to the simulated image patch to the characteristic information on the spatial plane to obtain the target image patch.

[0029] In a preferred embodiment, the step of processing the target retinal image based on the deep learning technology to obtain the lesion segmentation result includes:

[0030] Train the convolutional neural network with historical retinal images to obtain the trained learning model;

[0031] Input the target retinal image into the learning model to obtain the lesion segmentation result.

[0032] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0033] The present invention performs image enhancement by performing hierarchical processing on the retinal image, can perform local analysis and enhancement of the image in more detail and without disturbing, ensure the accuracy after image enhancement, ensure the details of image enhancement, can make the image contrast clearer, and facilitate the recognition and segmentation of the subsequent convolutional network. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings required in the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other accompanying drawings can also be obtained based on these drawings.

[0035] Figure 1 It is the flowchart of the method of the present invention. Specific embodiments

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0037] Embodiment 1, please refer to Figure 1 As shown, the method for segmenting diabetic retinopathy lesions based on deep learning in this embodiment includes the following steps:

[0038] S1. Determine a retinal image, divide the retinal image to obtain multiple image blocks, and perform hierarchical processing on the multiple image blocks to obtain multiple image slices;

[0039] S2. Perform image enhancement based on the image slices, and fuse the image slices after image enhancement to obtain a target retinal image;

[0040] S3. Process the target retinal image based on deep learning technology to obtain a lesion segmentation result.

[0041] In one embodiment, the step S1 of determining a retinal image, dividing the retinal image to obtain multiple image blocks, and performing hierarchical processing on the multiple image blocks to obtain multiple image slices includes:

[0042] S11. Obtain a retinal image that needs to be segmented for lesions, and divide the retinal image into multiple image blocks according to a preset size;

[0043] S12. Perform hierarchical processing on the multiple image blocks according to brightness respectively to obtain multiple image slices.

[0044] In one embodiment, the step S12 of performing hierarchical processing on the multiple image blocks according to brightness respectively to obtain multiple image slices includes:

[0045] S121. Configure a corresponding multiple of image beds for the multiple image blocks respectively, where an image bed includes multiple spatial planes;

[0046] S122. Define a hierarchical brightness value range, where the hierarchical brightness value range includes multiple preset hierarchical brightness ranges. Extract features from the image blocks according to the hierarchical brightness value range to obtain feature information corresponding to the multiple preset hierarchical brightness ranges;

[0047] S123. Place the multiple pieces of feature information one-to-one on the image bed respectively, and record the contours of the feature information on the image bed through the image bed to obtain multiple image slices.

[0048] In one embodiment, step S123 of placing the multiple pieces of feature information one-to-one on the image bed respectively and recording the contours of the feature information on the image bed through the image bed includes:

[0049] S1231. Place the multiple pieces of feature information one-to-one on the image bed in sequence according to the numerical sorting order of the hierarchical brightness ranges they belong to, and the image bed is set in sequence;

[0050] S1232. Set the brightness interval range of the corresponding feature information on multiple spatial planes of the image bed, extract and layer the feature information within the brightness interval range and place it on the multiple spatial planes, and mark the brightness of the feature information on the spatial planes. Among them, multiple photosensors are arranged on the multiple spatial planes, and the brightness value and the corresponding range of the feature information are recorded through the photosensors;

[0051] As described in the above steps S11 and S12, the retinal images to be segmented are determined. Before inputting the retinal images into the convolutional network for image lesion segmentation, the images need to be enhanced to better improve the accuracy and recognition of the images input into the convolutional network for segmentation. First, the retinal images are segmented into multiple image blocks. The segmentation rule is according to a preset size. For example, the retinal image is segmented according to a grid of A*A. Then, a corresponding number of image beds are configured for the multiple image blocks respectively. Among them, the image bed includes multiple spatial planes. Here, the spatial plane is a blank spatial layer, and this spatial layer is a blank database. This database exists in the form of a data layout as the spatial layer. The multiple spatial planes serve as the image bed. The image bed is composed of multiple spatial planes, that is, the database is divided again to obtain multiple spatial layers as spatial planes, which are used for subsequent storage and display of the image blocks after layering. A layered brightness value range is formulated. The layered brightness value range includes multiple preset layered brightness ranges. Feature extraction is performed on the image blocks according to the layered brightness value range. The feature extraction is the feature information that conforms to the layered brightness value range, and the feature information corresponding to multiple preset layered brightness ranges is obtained. Here, the feature information includes the feature contour and the corresponding specific brightness. Then, the multiple pieces of feature information are respectively placed one-to-one on the image bed. Here, the image bed is a general term for multiple spatial planes. After dividing according to the layered brightness value range, the feature information is further divided on the spatial plane according to the brightness interval range. Multiple photosensors are arranged on the spatial plane. Among them, the photosensors are evenly and closely distributed on the spatial plane, which is used to record the brightness value and the corresponding range of the corresponding feature information. The position of the feature information on the spatial plane partially corresponds to the original position on the image block, and has a relatively detailed division, which can analyze and enhance the image more carefully, making the details of the collected retinal images shown more clearly, and facilitating the subsequent output of more accurate segmentation information after the convolutional network.

[0052] In one embodiment, step S2 of performing image enhancement based on the image patches and fusing the image patches after image enhancement to obtain the target retinal image includes:

[0053] S21. Corresponding enhancement rules are respectively formulated for the multiple image blocks. Among them, the enhancement rule includes a brightness interval and a corresponding contrast enhancement difference value;

[0054] S22. The photosensors in the image patches are regulated based on the enhancement rules to obtain the simulated image patches;

[0055] S23. The simulated image patches that meet the preset conditions are used as the simulated image patches, and the brightness and the corresponding range of the photosensors corresponding to the simulated image patches are assigned to the feature information on the spatial plane to obtain the target image patches;

[0056] S24. Overlap multiple target image patches to obtain a target image block, and perform corresponding stitching among the multiple target image blocks according to their positions in the retinal image to obtain a target retinal image.

[0057] In one embodiment, step S22 of regulating the photosensors in the image patch based on the enhancement rule to obtain the simulated image patch includes:

[0058] S221. Obtain the brightness interval corresponding to the feature information and the existing contrast difference value as the adjustment base according to the brightness value and the corresponding range of the feature information recorded in the photosensor, and match the corresponding enhancement rule according to the adjustment base;

[0059] S222. Perform brightness simulation regulation on the photosensors in the image patch according to the matched enhancement rule, and load the brightness regulated by the photosensor simulation onto the image patch to obtain the simulated image patch.

[0060] In one embodiment, step S23 of using the simulated image patch that meets the preset conditions as the simulated image patch, and assigning the brightness and the corresponding range of the photosensor corresponding to the simulated image patch to the feature information on the spatial plane to obtain the target image patch includes:

[0061] S231. Set the preset conditions, where the preset condition is that the difference range of the contour coincidence between the simulated image patch and the image patch satisfies F;

[0062] S232. Use the simulated image patch that meets the preset conditions as the simulated image patch, and adjust the brightness and the corresponding range of the photosensor corresponding to the simulated image patch to the feature information on the spatial plane to obtain the target image patch.

[0063] As described in the above steps S21 - S24, the same enhancement rules are formulated for multiple image blocks (representing all image blocks). This means that although the image is enhanced after being divided locally, the enhancement rules used are the same. In this way, when all the subsequent image blocks are stitched together, the consistency of the adjusted image can be ensured. Here, the enhancement rules include the brightness interval and the corresponding contrast enhancement difference value. The brightness interval here represents the interval where the brightness of the enhanced feature information is located, and the contrast enhancement difference value represents the brightness to be adjusted within this brightness interval. For example, if the brightness range is represented by 0 to 1, where 0 represents black and 1 represents the brightest state, and the enhancement rule is that in the brightness interval of 0 - 0.2, the contrast enhancement difference value is the degree of darkening by 0.05. When the brightness of the feature information is 0.2, the contrast enhancement difference value is to darken by 0.05, and the brightness of the feature information will be adjusted to 0.15. In this way, the brightness in the image, the bright parts will become brighter and the shadow parts will become darker, enhancing the contrast of the entire image, which enables better identification of the detailed parts in the image. Then, according to the brightness values and corresponding ranges of the characteristic information recorded in the photosensors, the brightness intervals corresponding to the characteristic information and the existing contrast difference values are obtained as adjustment bases. According to the adjustment bases, the corresponding enhancement rules are matched, and the brightness values to be adjusted can be obtained. The brightness of the photosensors in the image patch is simulated and regulated according to the matched enhancement rules. The brightness loaded through the simulation of the photosensors is carried on the image patch, and the image patch after simulation is obtained. Since the characteristic information is set on the spatial plane, the photosensor is the unit for adjusting brightness and is a single cloud server. This cloud server is connected to the corresponding image patch. The image patch is the data layer. All the photosensors on a single spatial plane are interconnected. Multiple photosensors are responsible for the recording and regulation of the brightness in the corresponding ranges. All the storage ranges on the spatial plane are covered by the photosensors. In this way, after the characteristic information is laid on the spatial plane, it can be managed by the photosensors. Thus, by directly determining the brightness through the photosensors, the brightness of the responsible characteristic information can be adjusted. In an image patch, the characteristic information is not fully covered. Only the photosensors corresponding to the covered positions will have the operation of brightness adjustment. The simulation adjustment can be performed without directly being carried on the image patch. Since there is a corresponding relationship between the photosensors and the responsible characteristic information, the display of other ranges and brightness is the same. Therefore, the photosensors can be used to represent the image patch as the image patch after simulation. Then, preset conditions are set. The preset condition is that the difference range of the contour coincidence between the image patch after simulation and the image patch satisfies F. Since the photosensitive points are responsible for a certain range, there will be adjustments beyond the contour line during the adjustment. In this way, the difference range of the contour coincidence after the adjustment by the photosensors satisfying F represents normal. The image patch after simulation that meets the preset conditions is used as the simulated image patch. When it does not meet the conditions, the brightness of the photosensors in the affected range can be adjusted in detail by partitioning. For example, the adjustment range corresponding to the photosensor is Y. The Y range is divided according to the boundary of the brightness adjustment, which is equivalent to dividing the Y range into two parts for brightness adjustment. The brightness of the photosensors corresponding to the simulated image patch and the corresponding ranges are adjusted according to the characteristic information on the spatial plane to obtain the target image patch. The multiple target image patches are overlapped to obtain the target image block. The multiple target image blocks are correspondingly spliced according to their positions in the retinal image to obtain the target retinal image. When the characteristic information of multiple image patches is split, it is split corresponding to the positions. Therefore, they can be directly overlapped to obtain the target image block, enabling more detailed and non-intrusive local analysis and enhancement of the image, ensuring the accuracy of the image after enhancement, ensuring the details of the image enhancement, making the image contrast clearer, and facilitating the recognition and segmentation of the subsequent convolutional network.

[0064] In one embodiment, step S3 of processing the target retinal image based on deep learning technology to obtain the lesion segmentation result includes:

[0065] S31. Training a convolutional neural network with historical retinal images to obtain a trained learning model;

[0066] S32. Inputting the target retinal image into the learning model to obtain the lesion segmentation result.

[0067] As described in the above steps S31 and S32, the steps of training a convolutional neural network with historical retinal images to obtain a trained learning model are as follows: collect a large number of diabetic retinopathy images, which can be from ophthalmic examination equipment in hospitals, such as fundus cameras, optical coherence tomography (OCT), etc. Ensure that the images have different resolutions, lighting conditions, and lesion degrees to cover various possible situations. Label the lesion areas in the collected images, marking each pixel as belonging to the lesion or non-lesion category. Preprocess the images, including operations such as normalization, cropping, flipping, and rotation. Normalization can map the pixel values of the image to a certain range, which helps to accelerate the convergence speed of the model; cropping can remove the edge parts unrelated to the lesion in the image, reducing the data volume and computational amount; data augmentation operations such as flipping and rotation can increase the diversity of the data and prevent the model from overfitting. Input the preprocessed image data into the built FCN model, calculate the difference between the prediction result and the true label according to the loss function, and update the parameters of the network through the backpropagation algorithm to gradually reduce the loss function. During the training process, a validation set can be used to monitor the performance of the model. When the loss on the validation set no longer decreases or reaches the preset stop condition, stop the training to obtain the learning model. Then, input the target retinal image into the learning model to obtain the lesion segmentation result. The lesion segmentation result includes: Binary image data: This is the most direct segmentation result data. In a binary image, each pixel is usually represented by 0 and 1, where 0 represents the non-lesion area and 1 represents the lesion area. After converting the probability map to a binary image through a threshold, the pixel values in the image only have these two values, clearly distinguishing the lesion and non-lesion areas in the image, which is convenient for visually observing and analyzing the location and scope of the lesion. Lesion area statistical data: Some statistical data about the lesion area can be further calculated based on the binary image, such as the area, perimeter, diameter, etc. of the lesion. The lesion area refers to the sum of the number of pixel points marked as lesions in the image, which can reflect the size of the lesion; the perimeter is the length of the boundary of the lesion area, which is helpful for analyzing the shape and boundary characteristics of the lesion; the diameter can be obtained by calculating the diameter of the largest inscribed circle of the lesion area or other related methods, which can roughly describe the size range of the lesion. Lesion location information: The specific location information of the lesion in the retinal image can be obtained from the segmentation result. The central position of the lesion can be represented by coordinates, or the upper left and lower right coordinates of the lesion area can be recorded to determine the position range of the lesion in the image. In addition, by analyzing the relative position relationship between the lesion and other landmark structures in the image (such as blood vessels, optic nerve heads, etc.), the lesion can be more accurately located, providing more detailed information for subsequent diagnosis and treatment.

[0068] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims described above.

Claims

1. A method for segmenting diabetic retinopathy lesions based on deep learning, characterized in that Including the following steps: Determine a retinal image, divide the retinal image to obtain a plurality of image patches, and perform a layering process on the plurality of image patches to obtain a plurality of image slices; The steps of determining a retinal image, dividing the retinal image to obtain a plurality of image patches, and performing a layering process on the plurality of image patches to obtain a plurality of image slices include: Obtain a retinal image to be subjected to lesion segmentation, and divide the retinal image into a plurality of image patches according to a preset size; Perform a layering process on the plurality of image patches according to brightness respectively to obtain a plurality of image slices; The steps of performing a layering process on the plurality of image patches according to brightness respectively to obtain a plurality of image slices include: Correspondingly configure a plurality of corresponding image beds for the plurality of image patches, wherein an image bed includes a plurality of spatial planes; Formulate a layering brightness value range, the layering brightness value range includes a plurality of preset layering brightness ranges, extract features of the image patches according to the layering brightness value range to obtain feature information corresponding to the plurality of preset layering brightness ranges; One-to-one place the plurality of pieces of feature information on the image bed respectively, and record the contours of the feature information on the image bed through the image bed to obtain a plurality of image slices; Perform image enhancement based on the image slices, and fuse the image slices after image enhancement to obtain a target retinal image; The steps of performing image enhancement based on the image slices, and fusing the image slices after image enhancement to obtain a target retinal image include: Correspondingly formulate corresponding enhancement rules for the plurality of image patches, wherein the enhancement rules include a brightness interval and a corresponding contrast enhancement difference value; Regulate the photosensors in the image slices based on the enhancement rules to obtain simulated image slices; Use the simulated image slices that meet the preset conditions as simulated image slices, and assign the brightness and corresponding range of the photosensors corresponding to the simulated image slices to the feature information on the spatial plane to obtain target image slices; Overlap the plurality of target image slices to obtain target image patches, and splice the plurality of target image patches corresponding to their positions in the retinal image to obtain a target retinal image; Process the target retinal image based on deep learning technology to obtain a lesion segmentation result.

2. The method for segmenting diabetic retinopathy lesions based on deep learning according to claim 1, wherein: The steps of one-to-one place the plurality of pieces of feature information on the image bed respectively, and record the contours of the feature information on the image bed through the image bed include: One-to-one place the plurality of pieces of feature information on the image bed in sequence according to the numerical sorting of the layering brightness ranges where they are located, and the image bed is arranged in sequence; Set the brightness interval range of the corresponding feature information on the plurality of spatial planes of the image bed, extract and layer the feature information within the brightness interval range and place it on the plurality of spatial planes, and mark the brightness of the feature information on the spatial plane, wherein a plurality of photosensors are arranged on the plurality of spatial planes, and the brightness value and corresponding range of the feature information are recorded through the photosensors.

3. The method for segmenting diabetic retinopathy lesions based on deep learning according to claim 1, wherein: The steps of regulating the photosensors in the image slices based on the enhancement rules to obtain simulated image slices include: According to the brightness value and the corresponding range of the characteristic information recorded in the photosensor, obtain the brightness interval corresponding to the characteristic information and the existing contrast difference value as the adjustment base, and match the corresponding enhancement rule according to the adjustment base; Perform brightness simulation regulation on the photosensor in the image slice according to the matched enhancement rule, and load the brightness regulated by the photosensor simulation onto the image slice to obtain the image slice after simulation.

4. The method for segmenting diabetic retinopathy lesions based on deep learning according to claim 1, characterized in that: The step of using the image slice after simulation that meets the preset conditions as the simulated image slice, and assigning the brightness and the corresponding range of the photosensor corresponding to the simulated image slice to the characteristic information on the spatial plane to obtain the target image slice includes: Formulate a preset condition, where the preset condition is that the difference range of the contour coincidence between the image slice after simulation and the image slice satisfies F; Use the image slice after simulation that meets the preset conditions as the simulated image slice, and adjust the brightness and the corresponding range of the photosensor corresponding to the simulated image slice to the characteristic information on the spatial plane to obtain the target image slice.

5. The method for segmenting diabetic retinopathy lesions based on deep learning according to claim 1, wherein: The step of processing the target retinal image based on deep learning technology to obtain the lesion segmentation result includes: Train the convolutional neural network through historical retinal images to obtain the trained learning model; Input the target retinal image into the learning model to obtain the lesion segmentation result.

Images