A Lesion Region Segmentation System Based on Ultra-Wide-Angle Ophthalmic Images

Abstract

This invention provides a lesion region segmentation system based on ophthalmic ultra-wide-angle images, comprising: a generator training module for training a first lesion generator using a conventional color fundus image with pixel-level lesion annotations as the source domain and an unannotated ophthalmic ultra-wide-angle image as the target domain; a generator adjustment module for adding multiple convolutional modulation feature fusion blocks to the first lesion generator to form a second lesion generator; a first image analysis module for inputting the image to be detected into the second lesion generator to obtain a first segmented image of the lesion region with preliminary segmentation; and a second image analysis module for inputting the first segmented image into an adversarial domain discriminator to obtain a second segmented image of the lesion region with further segmentation. The beneficial effect is that this invention introduces convolutional modulation calculation into the segmentation network, reducing redundant information while obtaining sufficient contextual information, thus achieving higher segmentation accuracy.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a lesion region segmentation system based on ophthalmic ultra-wide-angle images. Background Technology

[0002] Ultra-wide-angle fundus (UWF) imaging is a novel, non-invasive imaging method. Compared to conventional color fundus images with a range of 30°-60°, UWF images can capture 180°-200° in a single image, covering approximately 80% of the fundus area. This allows for the visualization of more clinicopathological features associated with diabetic retinopathy. Extracting more lesion areas from UWF images plays a crucial role in promoting the screening and diagnosis of diabetic retinopathy. Clinical practice has shown that abnormal lesion areas present in UWF images, such as microaneurysms, hemorrhages, and exudates, are closely related to the severity of diabetic retinopathy. Accurately extracting these lesions and analyzing their morphology and type can shorten image evaluation time, reduce the incidence of failure to grade diabetic retinopathy, and provide new insights into the research of this disease.

[0003] For the reasons mentioned above, automatically and accurately extracting these abnormal lesion regions from fundus images is of great significance for the early screening and diagnosis of DR and for exploring the pathogenesis of the disease. To this end, researchers have proposed many deep convolutional networks based on conventional color fundus images and achieved significant results. However, these networks are all dedicated to learning highly representative features from labeled data to improve lesion segmentation performance. There is very little work on lesion segmentation based on UWF images because UWF image annotation takes a lot of time, which leads to a bottleneck in lesion segmentation performance, which cannot be improved further. In addition, there is a lot of redundant information, which will greatly affect the segmentation accuracy. Summary of the Invention

[0004] The problem to be solved by this invention is to provide a lesion region segmentation system based on ophthalmic ultra-wide-angle images, which makes full use of ophthalmic ultra-wide-angle images (UWF), breaks through the bottleneck of lesion segmentation of traditional conventional color fundus images, improves lesion segmentation performance and accuracy, and can reduce redundant information to improve segmentation accuracy.

[0005] To address the above problems, this invention provides a lesion region segmentation system based on ophthalmic ultra-wide-angle images, comprising:

[0006] A generator training module is used to train a first lesion generator by taking a regular color fundus image with pixel-level lesion annotations as the source domain and an unannotated ophthalmic ultra-wide-angle image as the target domain.

[0007] A generator adjustment module, connected to the generator training module, is used to add multiple convolutional modulation feature fusion blocks to the segmentation network structure in the first lesion generator to form a second lesion generator based on convolutional modulation.

[0008] A first image analysis module, connected to the generator adjustment module, is used to input the acquired image to be detected into the second lesion generator to obtain a first segmented image of the lesion region with preliminary segmentation;

[0009] A second image analysis module, connected to the first image analysis module, is used to input the first segmented image into a pre-constructed adversarial domain discriminator to obtain a second segmented image that further segments the lesion region, and output the second segmented image as the lesion region segmentation result.

[0010] Preferably, the first lesion generator is a U-shaped convolutional neural network.

[0011] Preferably, the U-shaped convolutional neural network includes an encoder and a decoder. The encoder is used to extract features from the current input image to obtain a corresponding first feature map, and the decoder is used to amplify the first feature map to obtain a corresponding second feature map.

[0012] Preferably, each of the convolutional modulation feature fusion blocks includes a plurality of first convolutional modulation feature fusion blocks and a second convolutional modulation feature fusion block. Each of the first convolutional modulation feature fusion blocks is located between adjacent convolutional layers performing upsampling operations within the U-shaped convolutional neural network. The second convolutional modulation feature fusion block is located between the last convolutional layer performing max pooling operations and the first convolutional layer performing upsampling operations.

[0013] Preferably, the first convolutional modulation feature fusion block includes:

[0014] The first processing layer is used to perform a 1×1 linear convolution operation on the first feature map extracted by the encoder to obtain the corresponding first convolution feature modulation value.

[0015] The second processing layer, which runs in parallel with the first processing layer, is used to sequentially perform a 1×1 linear convolution operation, GELU activation function processing, and a depthwise convolution operation with a kernel size of 11×11 and padding of 5 on the second feature map extracted by the decoder, to obtain the corresponding first similarity matrix.

[0016] The third processing layer, running above the first and second processing layers, is used to perform an Adama product on the first convolutional feature modulation value and the first similarity matrix to obtain the corresponding first cross-feature fusion map.

[0017] The fourth processing layer, running above the third processing layer, is used to perform pixel multiplication and addition operations on the first cross-feature fusion map, the first convolutional feature modulation value, and the first similarity matrix to obtain the input image of the next convolutional layer.

[0018] Preferably, the calculation formula for the input image of the next convolutional layer is as follows:

[0019] Z(X i )=A1⊙V1

[0020] A1=DConv k×k (W1X i )

[0021] V1 = W2X i

[0022] in,

[0023] Z(X i ) represents the input image of the next convolutional layer;

[0024] X i This represents the output image of the convolutional layer described above;

[0025] i represents the encoder-decoder index;

[0026] A1 represents the first similarity matrix;

[0027] V1 represents the modulation value of the first convolutional feature;

[0028] ⊙ represents Adama's product;

[0029] DConv k×k This represents a depthwise convolution operation with a kernel size of k×k;

[0030] k×k represents the kernel size of the depthwise convolution operation;

[0031] W1 represents the preset weight matrix of the first linear convolutional layer;

[0032] W2 represents the weight matrix of the preset first linear convolutional layer.

[0033] Preferably, the second convolutional modulation feature fusion block includes:

[0034] The fifth processing layer is used to perform a 1×1 linear convolution operation on the first feature map extracted by the encoder to obtain the corresponding second convolution feature modulation value;

[0035] The sixth processing layer, which runs in parallel with the fifth processing layer, is used to sequentially perform a 1×1 linear convolution operation, GELU activation function processing, and a depthwise convolution operation with a kernel size of 11×11 and padding of 5 on the first feature map to obtain the corresponding second similarity matrix.

[0036] The seventh processing layer, running above the fifth and sixth processing layers, is used to perform an Adama product on the second convolutional feature modulation value and the second similarity matrix to obtain the corresponding second cross-feature fusion map, which serves as the input image for the next convolutional layer.

[0037] Preferably, the calculation formula for the input image of the next convolutional layer is as follows:

[0038] Z(X i )=A2⊙V2

[0039] A2=DConv k×k (W1X i )

[0040] V2 = W2X i

[0041] in,

[0042] Z(X i ) represents the input image of the next convolutional layer;

[0043] X i This represents the output image of the convolutional layer described above;

[0044] i represents the encoder-decoder index;

[0045] A2 represents the second similarity matrix;

[0046] V2 represents the modulation value of the second convolution feature;

[0047] ⊙ represents Adama's product;

[0048] DConv k×k This represents a depthwise convolution operation with a kernel size of k×k;

[0049] k×k represents the kernel size of the depthwise convolution operation;

[0050] W1 represents the preset weight matrix of the first linear convolutional layer;

[0051] W2 represents the weight matrix of the preset first linear convolutional layer.

[0052] Preferably, the adversarial domain discriminator includes four stacked convolutional mapping tuples. The first convolutional mapping tuple is used to perform convolution operations with a kernel size of 7×7 and padding of 3. The second and third convolutional mapping tuples are used to perform convolution operations with a kernel size of 5×5 and padding of 2. The fourth convolutional mapping tuple is used to perform convolution operations with a kernel size of 3×2 and padding of 1.

[0053] Preferably, the stride of each of the convolutional mapping tuples is 2.

[0054] The present invention has the following beneficial effects:

[0055] 1) In this invention, a conventional color fundus image with pixel-level lesion annotations is used as the source domain, and an unannotated ophthalmic ultra-wide-angle image is used as the target domain. The first lesion generator is trained and used for lesion region segmentation, which breaks through the bottleneck of lesion segmentation of conventional color fundus images. Furthermore, based on the superior coverage of the fundus region by ultra-wide-angle images, the lesion segmentation performance and accuracy are improved.

[0056] 2) This invention introduces an adversarial domain discriminator and uses an adversarial domain transfer learning method to transfer rich feature knowledge from the source domain to the target domain, effectively alleviating the lack of pixel-level annotation in ultra-wide-angle images;

[0057] 3) In this invention, the convolutional modulation feature fusion block is introduced into the segmentation network structure, which reduces redundant information while obtaining sufficient contextual information, thereby achieving higher segmentation accuracy. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the structure of the present invention;

[0059] Figure 2 This is a schematic diagram of the structure of the second lesion generator and the adversarial domain discriminator of the present invention;

[0060] Figure 3 This is a schematic diagram of the structure of the convolutional modulation feature fusion block of the present invention;

[0061] Figure 4 This is a schematic diagram of the structure of the first convolutional modulation feature fusion block of the present invention;

[0062] Figure 5 This is a schematic diagram of the structure of the second convolutional modulation feature fusion block of the present invention;

[0063] Figure labeling: 1. Generator training module; 2. Generator adjustment module; 3. First image analysis module; 4. Second image analysis module; 5. First convolutional modulation feature fusion block; 51. First processing layer; 52. Second processing layer; 53. Third processing layer; 54. Fourth processing layer; 6. Second convolutional modulation feature fusion block; 61. Fifth processing layer; 62. Sixth processing layer; 63. Seventh processing layer. Detailed Implementation

[0064] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0065] In a preferred embodiment of the present invention, based on the aforementioned problems existing in the prior art, a lesion region segmentation system based on ophthalmic ultra-wide-angle images is provided, such as... Figure 1-3 As shown, it includes:

[0066] A generator training module 1 is used to train a first lesion generator by taking a regular color fundus image with pixel-level lesion annotations as the source domain and an unannotated ophthalmic ultra-wide-angle image as the target domain.

[0067] A generator adjustment module 2, connected to the generator training module 1, is used to add multiple convolutional modulation feature fusion blocks to the segmentation network structure in the first lesion generator to form a second lesion generator based on convolutional modulation.

[0068] A first image analysis module 3 is connected to a generator adjustment module 2, which is used to input the acquired image to be detected into a second lesion generator to obtain a first segmented image of the lesion region that is initially segmented.

[0069] A second image analysis module 4, connected to the first image analysis module 3, is used to input the first segmented image into a pre-constructed adversarial domain discriminator to obtain a second segmented image that further segments the lesion region, and output the second segmented image as the lesion region segmentation result.

[0070] Specifically, in this embodiment, based on a large amount of publicly available conventional color fundus image data resources and good model segmentation performance, a lesion region segmentation system is built. First, conventional color fundus images with pixel-level lesion annotations are used as the source domain, and unannotated UWF images are used as the target domain. With the incorporation of convolutional modulation feature fusion blocks, the two are used as inputs to train a second lesion generator G· based on convolutional modulation, so as to automatically obtain multiple lesion regions in the fundus image. This segmentation result is regarded as a structured output containing the feature similarity between the source domain and the target domain. In the output space, an adversarial domain discriminator D· is added to further complete the accurate segmentation of UWF image lesions. Finally, the extracted prior lesion knowledge is used to further assist in the screening and diagnosis of diabetic retinopathy.

[0071] Specifically, in this embodiment, when trained on a large amount of available source domain data, the lesion generator can learn useful information relevant to the task. However, due to data bias or domain shift between the source and target domain data, the segmentation model trained on the source domain cannot be directly applied to the target domain. To transfer rich source domain feature knowledge to the UWF feature space, this invention proposes adding a domain discriminator. This discriminator has a similar function to the discriminator in traditional GAN ​​networks. Traditional GAN ​​networks consist of a generator and a discriminator. The generator generates new data from random variables, and the discriminator distinguishes the difference between the generated samples and the real data. The ultimate goal is to make the data distribution of the generated samples converge as close as possible to the real data distribution. In this invention, the input is not a random variable following a certain distribution, but rather the source domain lesion segmentation result predicted by the lesion generator from pixel-level annotated data is used as the real data branch, and the target domain lesion segmentation result from unannotated pixel data is used as the fake sample branch. Adversarial learning promotes the learning of transferable features between the source and target domains, further reducing the differences between domains and achieving the goal of using source domain lesion segmentation maps to assist UWF in generating more refined lesion maps.

[0072] In a preferred embodiment of the present invention, the first lesion generator is a U-shaped convolutional neural network.

[0073] Specifically, in this embodiment, the U-shaped convolutional neural network used for segmentation mostly combines the high-resolution feature map of the encoder with the feature map of the decoder through skip connections, thereby introducing detailed information into the decoder. However, when processing peripheral fields and low-contrast pathological regions in UWF images, it is difficult to obtain sufficient information through the U-shaped convolutional neural network with its limited receptive field. This is mainly because the U-shaped convolutional neural network cannot model long-distance contextual interactions and spatial dependencies, resulting in a lack of sufficient information to handle complex medical image segmentation tasks. To alleviate this problem, this embodiment adds a new component, namely the convolutional modulation feature fusion block (CMFF), to the U-shaped convolutional neural network. This module can reduce redundant information while providing sufficient details to the decoder, thereby obtaining more accurate segmentation results.

[0074] In a preferred embodiment of the present invention, the U-shaped convolutional neural network includes an encoder and a decoder. The encoder is used to extract features from the current input image to obtain a corresponding first feature map, and the decoder is used to amplify the first feature map to obtain a corresponding second feature map.

[0075] Specifically, in this embodiment, Res-Unet is extended using a deeper base module and modified into a lesion generator. The encoder and decoder together consist of nine feature mapping tuples. ResNet50, pre-trained on ImageNet, is used as the base convolutional module. Except for the first tuple of the encoder, which inherits its first layer of convolutional kernel size of 3×3, BN layer, ReLU activation layer, and max pooling layer for feature extraction of the input image, the other four feature mapping tuples of the encoder are all composed of encoders of different depths in ResNet50.

[0076] In a preferred embodiment of the present invention, each convolutional modulation feature fusion block includes a plurality of first convolutional modulation feature fusion blocks 5 and a second convolutional modulation feature fusion block 6. Each first convolutional modulation feature fusion block 6 is located between adjacent convolutional layers performing upsampling operations within the U-shaped convolutional neural network. The second convolutional modulation feature fusion block 5 is located between the last convolutional layer performing max pooling operations and the first convolutional layer performing upsampling operations.

[0077] In a preferred embodiment of the present invention, such as Figure 4 As shown, the first convolutional modulation feature fusion block 5 includes:

[0078] The first processing layer 51 is used to perform a 1×1 linear convolution operation on the first feature map extracted by the encoder to obtain the corresponding first convolution feature modulation value.

[0079] The second processing layer 52, which runs in parallel with the first processing layer 51, is used to sequentially perform a 1×1 linear convolution operation, GELU activation function processing, and a depthwise convolution operation with a kernel size of 11×11 and padding of 5 on the second feature map extracted by the decoder, to obtain the corresponding first similarity matrix.

[0080] The third processing layer 53 runs above the first processing layer 51 and the second processing layer 52. It is used to perform the Hadamard product on the first convolutional feature modulation value and the first similarity matrix to obtain the corresponding first cross feature fusion map.

[0081] The fourth processing layer 54 runs above the third processing layer 53. It is used to perform pixel multiplication and addition operations on the first cross feature fusion map, the first convolution feature modulation value, and the first similarity matrix to obtain the input image of the next convolutional layer.

[0082] Specifically, in this embodiment, the first convolutional modulation feature fusion block 5 is... Figure 3 (a) Structure, in order to obtain sufficient contextual information, Figure 3 The structure shown in (a) replaces the traditional skip links to connect encoder-decoder structures at the same level, achieving full fusion of global information between encoder features. Specifically, when any encoder's first feature map X1 and decoder's second feature map X2 are used as inputs, a 1×1 linear convolution operation is first performed on the encoder's first feature map X1 to obtain its convolutional feature modulation value V1. Then, a 1×1 linear convolution operation is performed on the decoder's second feature map X2, followed by a GELU activation function and a depthwise convolution operation with a kernel size of 11×11 and padding of 5 to obtain the first similarity matrix 1 of the input. Finally, the Hadamard product of V1 and A1 is performed to obtain the cross-feature fusion map Z of the two inputs, which is then multiplied and added pixel by pixel with the first feature map X1 and the second feature map X2 to obtain the new input image of the next level.

[0083] In a preferred embodiment of the present invention, the calculation formula for the input image of the next convolutional layer is as follows:

[0084] Z(X i )=A1⊙V1

[0085] A1=DConv k×k (W1X i )

[0086] V1 = W2X i

[0087] in,

[0088] Z(X i () represents the input image for the next convolutional layer;

[0089] X i This represents the output image of the previous convolutional layer;

[0090] i represents the encoder-decoder index;

[0091] A1 represents the first similarity matrix;

[0092] V1 represents the modulation value of the first convolutional feature;

[0093] ⊙ represents Adama's product;

[0094] DConv k×k This represents a depthwise convolution operation with a kernel size of k×k;

[0095] k×k represents the kernel size of the depthwise convolution operation;

[0096] W1 represents the preset weight matrix of the first linear convolutional layer;

[0097] W2 represents the weight matrix of the preset first linear convolutional layer.

[0098] Specifically, in this embodiment, in the calculation of convolution modulation, the modulation value V of the depthwise convolution feature is utilized. i To simplify the correlation between different inputs, when given an input The output value Z can be calculated using a depthwise convolution with a kernel size of k×k and an Adama product.

[0099] Preferably, this calculation saves more storage space and makes all locations within the spatial dimension H×W related to pixels within a k×k square region centered at (h, w).

[0100] In a preferred embodiment of the present invention, such as Figure 5 As shown, the second convolutional modulation feature fusion block 6 includes:

[0101] The fifth processing layer 61 is used to perform a 1×1 linear convolution operation on the first feature map extracted by the encoder to obtain the corresponding second convolution feature modulation value;

[0102] The sixth processing layer 62, which runs in parallel with the fifth processing layer 61, is used to sequentially perform a 1×1 linear convolution operation, GELU activation function processing, and a depthwise convolution operation with a kernel size of 11×11 and padding of 5 on the first feature map to obtain the corresponding second similarity matrix.

[0103] The seventh processing layer 63, which runs above the fifth processing layer 61 and the sixth processing layer 62, is used to perform the Hadamard product on the second convolutional feature modulation value and the second similarity matrix to obtain the corresponding second cross feature fusion map, which serves as the input image for the next convolutional layer.

[0104] Specifically, in this embodiment, the second convolutional modulation feature fusion block 6 is... Figure 3 (b) Structure, Figure 3 (b) has a single input structure, and the solution process is similar to... Figure 3 Similar to (a), it is introduced at the last encoder to learn remote knowledge from the input image and associate each pixel in the high-level semantic features learned by the encoder to obtain the receptive field of the entire image, thereby achieving accurate segmentation of key UWF lesions.

[0105] In a preferred embodiment of the present invention, the calculation formula for the input image of the next convolutional layer is as follows:

[0106] Z(X i )=A2⊙V2

[0107] A2=DConv k×k (W1X i )

[0108] V2 = W2X i

[0109] in,

[0110] Z(X i () represents the input image for the next convolutional layer;

[0111] X i This represents the output image of the previous convolutional layer;

[0112] i represents the index of the encoder-decoder;

[0113] A2 represents the second similarity matrix;

[0114] V2 represents the modulation value of the second convolution feature;

[0115] ⊙ represents Adama's product;

[0116] DConv k×k This represents a depthwise convolution operation with a kernel size of k×k;

[0117] k×k represents the kernel size of the depthwise convolution operation;

[0118] W1 represents the preset weight matrix of the first linear convolutional layer;

[0119] W2 represents the weight matrix of the preset first linear convolutional layer.

[0120] In a preferred embodiment of the present invention, the adversarial domain discriminator includes four stacked convolutional mapping tuples. The first convolutional mapping tuple is used to perform convolution operations with a kernel size of 7×7 and padding of 3. The second and third convolutional mapping tuples are used to perform convolution operations with a kernel size of 5×5 and padding of 2. The fourth convolutional mapping tuple is used to perform convolution operations with a kernel size of 3×2 and padding of 1.

[0121] Specifically, in this embodiment, the output of the last convolutional mapping tuple is passed through an adaptive average pooling layer to reduce the spatial dimension of the feature map, and then through a fully connected layer and a sigmoid activation function to produce a binary output, which is used to distinguish whether the lesion map predicted by the output comes from the source domain or the target domain.

[0122] In a preferred embodiment of the present invention, the stride of each convolutional mapping tuple is 2.

[0123] While the disclosure is as stated above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of this disclosure, and all such changes and modifications will fall within the protection scope of this invention.

Claims

1. A lesion region segmentation system based on ultra-wide-angle ophthalmic images, characterized in that, include: A generator training module (1) is used to train a first lesion generator by taking a regular color fundus image with pixel-level lesion annotations as the source domain and an unannotated ophthalmic ultra-wide-angle image as the target domain. The first lesion generator is a U-shaped convolutional neural network. A generator adjustment module (2) is connected to the generator training module (1) and is used to add multiple convolutional modulation feature fusion blocks to the segmentation network structure in the first lesion generator to form a second lesion generator based on convolutional modulation. A first image analysis module (3) is connected to the generator adjustment module (2) and is used to input the acquired image to be detected into the second lesion generator to obtain a first segmented image of the lesion region that is initially segmented; A second image analysis module (4) is connected to the first image analysis module (3) and is used to input the first segmented image into a pre-constructed adversarial domain discriminator to obtain a second segmented image that further segments the lesion region, and output the second segmented image as the lesion region segmentation result. The first lesion generator includes an encoder and a decoder. The encoder is used to extract features from the current input image to obtain a corresponding first feature map, and the decoder is used to enlarge the first feature map to obtain a corresponding second feature map. Each of the convolutional modulation feature fusion blocks includes multiple first convolutional modulation feature fusion blocks (5) and one second convolutional modulation feature fusion block (6). Each first convolutional modulation feature fusion block (5) is located between adjacent convolutional layers performing upsampling operations within the U-shaped convolutional neural network. The second convolutional modulation feature fusion block (6) is located between the last convolutional layer performing max pooling operations and the first convolutional layer performing upsampling operations. The first convolutional modulation feature fusion block (5) includes: The first processing layer (51) is used to perform a 1×1 linear convolution operation on the first feature map extracted by the encoder to obtain the corresponding first convolution feature modulation value. The second processing layer (52) runs in parallel with the first processing layer (51) and is used to sequentially perform 1×1 linear convolution, GELU activation function processing, and depthwise convolution with kernel size of 11×11 and padding of 5 on the second feature map extracted by the decoder to obtain the corresponding first similarity matrix. The third processing layer (53) runs above the first processing layer (51) and the second processing layer (52) and is used to perform the Adama product on the first convolutional feature modulation value and the first similarity matrix to obtain the corresponding first cross feature fusion map. The fourth processing layer (54) runs above the third processing layer (53) and is used to perform pixel multiplication and addition operations on the first cross feature fusion map, the first convolution feature modulation value, and the first similarity matrix to obtain the input image of the next convolutional layer. The second convolutional modulation feature fusion block (6) includes: The fifth processing layer (61) is used to perform a 1×1 linear convolution operation on the first feature map extracted by the encoder to obtain the corresponding second convolution feature modulation value; The sixth processing layer (62) runs in parallel with the fifth processing layer (61) and is used to sequentially perform 1×1 linear convolution, GELU activation function processing, and depthwise convolution with kernel size of 11×11 and padding of 5 on the first feature map to obtain the corresponding second similarity matrix. The seventh processing layer (63) runs above the fifth processing layer (61) and the sixth processing layer (62) and is used to perform an Adama product on the second convolutional feature modulation value and the second similarity matrix to obtain the corresponding second cross feature fusion map, which serves as the input image for the next convolutional layer.

2. The lesion area segmentation system according to claim 1, characterized in that, The calculation formula for the first cross-feature fusion map in the third processing layer (53) is as follows: ； in, This represents the first cross-feature fusion map; This represents the first feature map; Indicates the encoder-decoder index; This represents the first similarity matrix; This represents the modulation value of the first convolutional feature; Represents the Adama product; Indicates the kernel size as Depth-wise convolution operations; This indicates the kernel size for depthwise convolution operations; This represents the second feature map; This represents the preset first weight matrix; This represents the preset second weight matrix.

3. The lesion area segmentation system according to claim 1, characterized in that, The calculation formula for the second cross-feature fusion map in the seventh processing layer (63) is as follows: ； in, This represents the second cross-feature fusion map; This represents the first feature map; Indicates the encoder-decoder index; This represents the second similarity matrix; This represents the second convolutional feature modulation value; Represents the Adama product; Indicates the kernel size as Depth-wise convolution operations; This indicates the kernel size for depthwise convolution operations; This represents the preset first weight matrix; This represents the preset second weight matrix.

4. The lesion area segmentation system according to claim 1, characterized in that, The adversarial domain discriminator includes four stacked convolutional mapping tuples. The first convolutional mapping tuple is used to perform a convolution operation with a kernel size of 7×7 and padding of 3. The second and third convolutional mapping tuples are used to perform a convolution operation with a kernel size of 5×5 and padding of 2. The fourth convolutional mapping tuple is used to perform a convolution operation with a kernel size of 3×3 and padding of 1.

5. The lesion area segmentation system according to claim 4, characterized in that, The stride of each of the convolutional mapping tuples is 2.

Images