Medical image segmentation method based on gating aggregation and Inception

Abstract

The invention provides a medical image segmentation method based on gating aggregation and Inception, and belongs to the technical field of medical image segmentation of image segmentation. The technical problem that a traditional method does not fully utilize an encoder to extract features and fuse the features is solved. According to the technical scheme, the method comprises the following steps of 1, preprocessing an input skin disease medical image; step 2, a Unet network based on gating aggregation and Inception is constructed; step 3, inputting the preprocessed skin disease medical image data set into a Unet network based on gating aggregation and Inception for training to obtain an optimal model; and step 4, after training is completed, inputting a skin disease medical image test set into the obtained optimal model, and detecting a segmentation result of a focus in the skin disease medical image. The method has the advantage that high segmentation accuracy and robustness are guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing methods, and in particular to a medical image segmentation method based on a Unet network of gated aggregation and Inception. Background Art

[0002] Medical image segmentation technology aims to extract target regions of interest from complex medical images and is widely used in biomedical research and clinical diagnosis. Its main application areas include disease detection and diagnosis, surgical planning, treatment effect evaluation, and medical research. Through precise segmentation, researchers and doctors can perform quantitative analysis on specific physiological or pathological structures, thereby improving the accuracy of diagnosis and the effectiveness of treatment.

[0003] Image segmentation algorithms can be divided into traditional methods and deep learning-based methods. Traditional algorithms include methods based on threshold, edge detection, and region growing. Although they work well in some simple scenarios, they have limited performance in complex medical images. In contrast, deep learning-based segmentation algorithms have the advantages of automatic feature extraction and high precision, especially convolutional neural networks (CNNs) and their derivative models, such as U-shaped structures (UNet), which perform well in medical image segmentation and significantly improve segmentation effects and efficiency.

[0004] The application of deep learning in medical image segmentation started early, and it has developed rapidly with the continuous advancement of algorithms and computing power. Early research focused on the segmentation of organs and lesions using the UNet architecture, and gradually expanded to the processing of multimodal images and complex lesions. Deep learning methods can capture subtle features in images through training with a large amount of annotated data, significantly improving the accuracy and robustness of segmentation and promoting the advancement of medical image analysis.

[0005] Although existing image segmentation networks have achieved remarkable results in many tasks, they still have some shortcomings. First, the fixed-size convolution kernel limits the full extraction of feature information; second, it ignores the intra-class differences in medical images, resulting in feature redundancy; in addition, for lesion images with irregular morphology, existing models have difficulty in accurately segmenting their contours. To address these problems, researchers have proposed a variety of improvement schemes, such as introducing deformable convolution, attention mechanism, and multi-scale feature fusion, to improve the performance of segmentation networks.

[0006] At present, medical image segmentation models urgently need to be further optimized to cope with complex and changing clinical needs. Specific improvement directions include enhancing the model's adaptability to different lesion morphologies, improving the processing effect of low-contrast and noisy images, and reducing dependence on large amounts of labeled data. At the same time, combining multimodal information and introducing prior knowledge are also important ways to improve segmentation accuracy. Through continuous algorithm innovation and model optimization, medical image segmentation technology will provide more reliable and efficient support for clinical diagnosis and treatment.

[0007] How to solve the above technical problems is the subject faced by the present invention. Summary of the invention

[0008] The technical problem to be solved by the present invention is how to achieve a segmentation method of a Unet network with higher detection accuracy while maintaining a low parameter amount and a fast detection speed, and to provide a medical image segmentation method based on gated aggregation and Inception.

[0009] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: a medical image segmentation method based on gated aggregation and Inception, comprising the following steps:

[0010] Step 1: Preprocess the input dermatological medical images, mainly image denoising and image enhancement, and divide the dataset into training set and test set according to the ratio;

[0011] Step 2: Constructing a Unet network based on gated aggregation and Inception. The Unet network based on gated aggregation and Inception uses an Inception-based module to extract multi-scale features and enrich lesion feature information. Then, through gated aggregation, the selectivity of the network in the feature aggregation process is enhanced, and the network is allowed to better control which feature information is aggregated.

[0012] Step 3: Input the preprocessed dermatology medical image dataset into the Unet network based on gated aggregation and Inception for training to obtain the optimal model;

[0013] Step 4: After the training is completed, the dermatology medical image test set is input into the obtained optimal model to detect the segmentation results of the lesions in the dermatology medical images.

[0014] The specific process in step 1 is as follows:

[0015] Step 1.1: Read the color image through the image processing library and obtain its RGB channel information. RGB images are composed of three color channels: red, green, and blue, and each channel contains corresponding pixel values.

[0016] Step 1.2: Normalize the pixel values ​​of the image to the range [0,1]. The normalization formula is as follows:

[0017]

[0018] Step 1.3: Bilateral filtering denoising.

[0019] First, the spatial distance weight is calculated to take into account the geometric distance between pixels. The spatial distance weight calculation formula is as follows:

[0020]

[0021] Next, the weights between pixel values ​​are calculated to preserve edge information. The formula for calculating pixel value weights is as follows:

[0022]

[0023] Then, the spatial and pixel value weights are combined to obtain the normalized weights, which are calculated as follows:

[0024]

[0025] Finally, the normalized weights are used to perform weighted averaging on the pixels in the neighborhood to obtain the filtered image. The formula for the filtered pixel value is as follows:

[0026]

[0027] Step 1.4: Image enhancement based on histogram equalization.

[0028] First, the denoised RGB image is converted to the YCrCb color space to separate the luminance and chrominance information.

[0029] Subsequently, the brightness channel Y is extracted from the YCrCb image and its histogram is processed separately to enhance the brightness contrast of the image.

[0030] Next, adaptive histogram equalization (such as CLAHE) is applied to process the brightness channel Y to enhance local contrast and improve image details.

[0031] Then, the equalized luminance channel Y is recombined with the original chrominance channels Cr and Cb to form an enhanced YCrCb image.

[0032] Finally, the processed YCrCb image is converted back to RGB color space to obtain an enhanced image with better visual effect.

[0033] Step 1.5: Divide the skin disease image dataset after data enhancement in step 1.4 into a training set and a test set in a ratio of 8:2.

[0034] The specific process in step 2 is as follows:

[0035] Step 2.1: Use the improved Inception module to extract multi-scale feature information. The improved Inception module consists of four parallel branches. Each branch independently completes its own feature extraction process.

[0036]

[0037] Step 2.2: Branch 1.

[0038] The number of channels is compressed through 1x1 convolution, and then the spatial features are extracted through 3x3 convolution. The detail features in the horizontal and vertical directions are extracted through 3x1 and 1x3 convolution respectively. Finally, the depth-wise separable convolution is used to further reduce the amount of calculation and enhance the feature expression capability.

[0039] X 1，1 =Conv 1×1 (X)

[0040] X 1,2 =Conv 3×3 (X 1，1 )

[0041] X 1,3 =Conv 3×1 (X 1,2 )+Conv 1×3 (X 1,2 )

[0042] X 1,4 =DepthwiseSeparableConv(X 1,3 )

[0043] Step 2.3: Branch 2.

[0044] Use 1x1 convolution to reduce the number of channels, then use 3x1 and 1x3 convolution to extract features in the horizontal and vertical directions respectively, and finally use the attention mechanism (like SE module) to enhance important features and suppress irrelevant features.

[0045] X 2，1 =Conv 1×1 (X)

[0046] X 2,2 =Conv 3×1 (X 2，1 )+Conv 1×3 (X 2，1 )

[0047] X 2,3 =Attention(X2,2 )

[0048] Step 2.4: Branch 3.

[0049] First, global context information is captured through pooling (average pooling or maximum pooling), then the feature channels are further compressed through 1x1 convolution, and finally the nonlinear ability of feature expression is enhanced through bilinear pooling.

[0050] X 3，1 =Pool(X)

[0051] X 3,2 =Conv 1×1 (X 3，1 )

[0052] X 3,3 =BilinearPooling(X 3,2 )

[0053] Step 2.5: Branch 4.

[0054] The original feature information is preserved through 1x1 convolution, and a larger receptive field is introduced through extended convolution (like 5x5 convolution) to capture a wider range of contextual information.

[0055] X 4，1 =Conv 1×1 (X)

[0056] X 4，2 =DilatedConv(X 4，1 )

[0057] Step 2.6: Branch splicing.

[0058] The outputs of branches 1, 2, 3, and 4 are concatenated in the channel dimension to form a comprehensive feature map. The concatenated feature map is then batch normalized and ReLU activated to normalize and activate the features.

[0059] X concat =Concat(X 1,4 ,X 2,3 ,X 3,3 ,X 4,2 )

[0060] X bn =BatchNorm(X concat )

[0061] X final =ReLU(X bn )

[0062] Step 2.7: Use the gated aggregation mechanism to perform weighted fusion on the concatenated feature maps obtained in step 2.6. First, use global average pooling to extract global information, then use the fully connected layer to generate weight coefficients, and then use the Sigmoid activation function to ensure that the weights are in the range of [0,1]. Finally, perform weighted fusion and adjust the weights of each branch feature by channel-by-channel multiplication.

[0063] G = GAP(X final )

[0064] G reduced =ReLU(Conv 1×1 (G)

[0065] G restored =σ(Conv 1×1 (G reduced ))

[0066] X gated =X final ⊙G restored

[0067] Y final =ReLU(X gated )

[0068] The specific process in step 3 is as follows:

[0069] Step 3.1: Initialize the parameters of the Unet network based on gated aggregation and Inception;

[0070] Step 3.2: Set the training parameters, the training epoch is 120, the single training data size (Batch Size) is 4, the initial learning rate (lr) of the network is set to 0.0001, and the Adam optimizer is used;

[0071] Step 3.3: Training uses DSC, mIoU and Accuracy as evaluation indicators of the model to evaluate the training results. The function expression of the DSC coefficient is as follows:

[0072]

[0073] Among them, TP, FP, and FN represent true positives, false positives, and false negatives, respectively.

[0074] The function expression of the mIoU coefficient is as follows:

[0075]

[0076] Among them, TP, FP, and FN represent true positives, false positives, and false negatives, respectively.

[0077] The function expression of the Accuracy coefficient is as follows:

[0078]

[0079] Among them, TP, FP, FN, and TN represent true positives, false positives, false negatives, and true negatives, respectively.

[0080] Step 3.4: Use the improved UNet network model for iterative training to extract the characteristic information of skin disease images and obtain the optimal model for segmentation result output.

[0081] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0082] 1. In the image preprocessing stage, an image enhancement method based on histogram equalization is adopted, especially for local contrast enhancement of the brightness channel. By converting the image from RGB color space to YCrCb color space, the brightness and chromaticity information are separated, and the details and contrast of the local area are improved through adaptive histogram equalization. Compared with the traditional global histogram equalization method, this adaptive enhancement strategy can improve the visual effect of the image more finely, especially when processing complex skin disease images, which helps to improve the accuracy of the subsequent segmentation network.

[0083] 2. The method of the present invention proposes an improved Inception module, which realizes multi-scale feature extraction through four parallel branches. Each branch independently completes feature extraction at different levels and directions, so as to effectively capture image features of different scales and details. Specifically, branch 1 uses depthwise separable convolution, branch 2 introduces the attention mechanism SE module, branch 3 uses bilinear pooling to enhance nonlinear expression, and branch 4 increases the receptive field by expanding convolution. This multi-scale and multi-directional feature extraction can fully mine the detailed information of the image and improve the accuracy and robustness of medical image segmentation.

[0084] 3. In traditional feature fusion methods, feature maps are usually simply concatenated or weighted averaged, while the gated aggregation mechanism introduces global average pooling and weight coefficients generated by the fully connected layer to perform weighted fusion of features from different branches, and can dynamically adjust the importance of each branch feature according to different regions and contextual information of the image. By using the Sigmoid activation function to ensure that the weight is between [0,1], this adaptive weighting mechanism can more effectively aggregate useful information, suppress irrelevant features, and further improve the image segmentation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0086] Figure 1 It is a schematic diagram of the overall process of an embodiment of the present invention.

[0087] Figure 2 The figure is a schematic diagram of the Unet network structure based on gated aggregation and Inception in the method described in an embodiment of the present invention.

[0088] Figure 3 Schematic diagram of the structure of the improved Inception in an embodiment of the present invention.

[0089] Figure 4 Schematic diagram of the structure of the gated aggregation used in the embodiment of the present invention.

[0090] Figure 5 Schematic diagram of the structure of the depthwise separable convolution used in an embodiment of the present invention.

[0091] Figure 6 Schematic diagram of the structure of the attention mechanism (SE module) used in the improved Inception module in an embodiment of the present invention. DETAILED DESCRIPTION

[0092] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0093] Example 1

[0094] Figure 1 The overall process diagram of an embodiment of the present invention is shown. First, a skin disease image dataset is constructed; then, the images in the dataset are preprocessed and divided into a training set and a test set according to a certain ratio. Next, a medical image segmentation model based on the Unet network of gated aggregation and Inception is constructed. Then, the preprocessed training set is input into the improved UNet model for training to obtain the optimal model. After the training is completed, the test set is input into the optimal model to detect the segmentation results of the lesions in the skin disease images.

[0095] The specific implementation steps of the above method are as follows:

[0096] Step 1.1: Read the color image through the image processing library and obtain its RGB channel information. RGB images are composed of three color channels: red, green, and blue, and each channel contains corresponding pixel values.

[0097] Step 1.2: Normalize the pixel values ​​of the image to the range [0,1]. The normalization formula is as follows:

[0098]

[0099] Step 1.3: Bilateral filtering denoising.

[0100] First, the spatial distance weight is calculated to take into account the geometric distance between pixels. The spatial distance weight calculation formula is as follows:

[0101]

[0102] Next, the weights between pixel values ​​are calculated to preserve edge information. The formula for calculating pixel value weights is as follows:

[0103]

[0104] Then, the spatial and pixel value weights are combined to obtain the normalized weights. The normalized weight calculation formula is as follows:

[0105]

[0106] Finally, the normalized weights are used to perform weighted averaging on the pixels in the neighborhood to obtain the filtered image. The formula for the filtered pixel value is as follows:

[0107]

[0108] Step 1.4: Image enhancement based on histogram equalization.

[0109] First, the denoised RGB image is converted to the YCrCb color space to separate the luminance and chrominance information.

[0110] Subsequently, the brightness channel Y is extracted from the YCrCb image and its histogram is processed separately to enhance the brightness contrast of the image.

[0111] Next, adaptive histogram equalization (such as CLAHE) is applied to process the brightness channel Y to enhance local contrast and improve image details.

[0112] Then, the equalized luminance channel Y is recombined with the original chrominance channels Cr and Cb to form an enhanced YCrCb image.

[0113] Finally, the processed YCrCb image is converted back to RGB color space to obtain an enhanced image with better visual effect.

[0114] Step 1.5: Divide the skin disease image dataset after data enhancement in step 1.4 into a training set and a test set in a ratio of 8:2.

[0115] Step 2.1: Use the improved Inception module to extract multi-scale feature information. The improved Inception module consists of four parallel branches. Each branch independently completes its own feature extraction process.

[0116]

[0117] Step 2.2: Branch 1.

[0118] The number of channels is compressed through 1x1 convolution, and then the spatial features are extracted through 3x3 convolution. The detail features in the horizontal and vertical directions are extracted through 3x1 and 1x3 convolution respectively. Finally, the depth-wise separable convolution is used to further reduce the amount of calculation and enhance the feature expression capability.

[0119] X 1，1 =Conv 1×1 (X)

[0120] X 1,2 =Conv 3×3 (X 1，1 )

[0121] X 1,3 =Conv 3×1 (X 1,2 )+Conv 1×3 (X 1,2 )

[0122] X 1,4 =DepthwiseSeparableConv(X 1,3 )

[0123] Step 2.3: Branch 2.

[0124] Use 1x1 convolution to reduce the number of channels, then use 3x1 and 1x3 convolution to extract features in the horizontal and vertical directions respectively, and finally use the attention mechanism (such as SE module) to enhance important features and suppress irrelevant features.

[0125] X 2，1 =Conv 1×1 (X)

[0126] X 2,2 =Conv 3×1 (X 2，1 )+Conv 1×3 (X 2，1 )

[0127] X 2,3 =Attention(X 2,2 )

[0128] Step 2.4: Branch 3.

[0129] First, global context information is captured through pooling (average pooling or maximum pooling), then the feature channels are further compressed through 1x1 convolution, and finally the nonlinear ability of feature expression is enhanced through bilinear pooling.

[0130] X 3，1 =Pool(X)

[0131] X 3,2 =Conv 1×1 (X 3，1 )

[0132] X 3,3 =BilinearPooling(X 3,2 )

[0133] Step 2.5: Branch 4.

[0134] The original feature information is retained through 1x1 convolution, and a larger receptive field is introduced through extended convolution (such as 5x5 convolution) to capture a wider range of contextual information.

[0135] X 4，1 =Conv 1×1 (X)

[0136] X 4，2 =DilatedConv(X 4，1 )

[0137] Step 2.6: Branch splicing.

[0138] The outputs of branches 1, 2, 3, and 4 are concatenated in the channel dimension to form a comprehensive feature map. The concatenated feature map is then batch normalized and ReLU activated to normalize and activate the features.

[0139] X concat =Concat(X 1,4 ,X 2,3 ,X 3,3 ,X 4,2 )

[0140] X bn =BatchNorm(X concat )

[0141] X final =ReLU(X bn )

[0142] Step 2.7: Use the gated aggregation mechanism to perform weighted fusion on the concatenated feature maps obtained in step 2.6. First, use global average pooling to extract global information, then use the fully connected layer to generate weight coefficients, and then use the Sigmoid activation function to ensure that the weights are in the range of [0,1]. Finally, perform weighted fusion and adjust the weights of each branch feature by channel-by-channel multiplication.

[0143] G = GAP(X final )

[0144] G reduced =ReLU(Conv 1×1 (G)

[0145] G restored =σ(Conv 1×1 (G reduced ))

[0146] X gated =X final ⊙G restored

[0147] Y final =ReLU(X gated )

[0148] Step 3.1: Initialize the parameters of the Unet network based on gated aggregation and Inception;

[0149] Step 3.2: Set the training parameters, the training epoch is 120, the single training data size (Batch Size) is 4, the initial learning rate (lr) of the network is set to 0.0001, and the Adam optimizer is used;

[0150] Step 3.3: Training uses DSC, mIoU and Accuracy as evaluation indicators of the model to evaluate the training results. The function expression of the DSC coefficient is as follows:

[0151]

[0152] Among them, TP, FP, and FN represent true positives, false positives, and false negatives, respectively.

[0153] The function expression of the mIoU coefficient is as follows:

[0154]

[0155] Among them, TP, FP, and FN represent true positives, false positives, and false negatives, respectively.

[0156] The function expression of the Accuracy coefficient is as follows:

[0157]

[0158] Among them, TP, FP, FN, and TN represent true positives, false positives, false negatives, and true negatives, respectively.

[0159] Step 3.4: Use the improved UNet network model for iterative training to extract the characteristic information of skin disease images and obtain the optimal model for segmentation result output.

[0160] Step 4: After the training is completed, the dermatology medical image test set is input into the obtained optimal model to detect the segmentation results of the lesions in the dermatology medical images.

[0161] The results of comparing the proposed method with the current mainstream skin disease image segmentation model on the ISIC 2018 dataset are shown in Table 1.

[0162] Table 1 Performance comparison of segmentation models on the ISIC 2018 dataset

[0163]

[0164]

[0165] Table 1 compares the performance of different segmentation models on the ISIC-2018 dataset, and the evaluation indicators include Dice similarity coefficient (DSC), mean intersection over union (mIoU) and accuracy (Acc). From the results, the proposed method achieves the best performance in all three indicators, with DSC of 88.40%, mIoU of 79.21% and Acc of 94.34%, indicating that its segmentation performance is better than other comparison models.

[0166] Specifically, the DSC, mIoU and Acc of the traditional UNet are 87.35%, 77.43% and 93.45% respectively, which is relatively low. UNet++ has achieved certain improvements through structural improvements, with DSC and mIoU reaching 87.46% and 78.13% respectively. Attention-UNet introduces the attention mechanism, with DSC of 87.67% and mIoU of 78.26%, showing an enhanced effect on the target area. SANet performs relatively well in various indicators, with DSC and mIoU reaching 88.23% and 78.96% respectively, which is only slightly lower than the method of the present invention.

[0167] In summary, the proposed method has achieved significant performance improvement in the skin disease image segmentation task through effective improvement strategies, especially in the DSC and mIoU indicators, showing stronger similarity and overlap between the predicted area and the true area. This proves the effectiveness and superiority of the proposed method in processing the ISIC-2018 dataset.

[0168] Example 2

[0169] This example further verifies the effectiveness of the method of the present invention in skin disease image segmentation by comparing it with the traditional FCN (full convolutional network) and DeepLabV3 models. The experiment uses the same ISIC 2018 dataset as in Example 1, and the evaluation indicators include Dice similarity coefficient (DSC), mean intersection over union (mIoU) and accuracy (Acc).

[0170] Table 2 Performance comparison of different segmentation models on the ISIC 2018 dataset (Example 2)

[0171] Segmentation Model DSC QUR Acc FCN 84.50 73.12 91.25 DeepLabV3 86.34 75.62 92.61 Method of the present invention 88.40 79.21 94.34

[0172] In comparison with FCN and DeepLabV3, the method of the present invention shows significant advantages:

[0173] 1. DSC: The DSC of the proposed method is 88.40%, which is 3.90% higher than FCN and 2.06% higher than DeepLabV3. This shows that the proposed method can better segment the target area in the skin disease image and reduce the probability of mis-segmentation.

[0174] 2. mIoU: The mIoU of the proposed method is 79.21%, which is higher than 73.12% of FCN and 75.62% of DeepLabV3, showing the better performance of the proposed method in accurately locating the target area.

[0175] 3. Acc: The accuracy of the proposed method is 94.34%, which is significantly improved compared to 91.25% of FCN and 92.61% of DeepLabV3, reflecting a higher overall accuracy.

[0176] This set of experimental results shows that the method of the present invention is superior to FCN and DeepLabV3 in terms of DSC, mIoU and Acc indicators, indicating that the technical solution of the present invention has higher accuracy and robustness when processing skin disease image segmentation tasks.

[0177] Example 3

[0178] This example further compares the method of the present invention with other popular segmentation network models, such as SegNet and PSPNet, in order to verify the superiority of the method of the present invention in the task of skin disease image segmentation.

[0179] Table 2 Performance comparison of different segmentation models on the ISIC 2018 dataset (Example 3)

[0180] Segmentation Model DSC QUR Acc SegNet 82.74 71.33 89.45 PSPNet 85.98 74.42 91.76 Method of the present invention 88.40 79.21 94.34

[0181] Compared with SegNet and PSPNet, the proposed method shows obvious advantages:

[0182] 1. DSC: The DSC of the proposed method is 88.40%, which is 5.66% higher than SegNet and 2.42% higher than PSPNet. This shows that the proposed method has significantly improved the overlap and similarity of the segmented regions.

[0183] 2. mIoU: The mIoU of the proposed method is 79.21%, which is better than 71.33% of SegNet and 74.42% of PSPNet, indicating that the proposed method can more accurately segment the detail areas of skin disease images.

[0184] 3. Acc: The accuracy of the proposed method is 94.34%, which is significantly improved compared with 89.45% of SegNet and 91.76% of PSPNet, proving the advantage of the proposed method in overall prediction accuracy.

[0185] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A medical image segmentation method based on gated aggregation and Inception, characterized in that: The following steps are involved: Step 1: Preprocess the input dermatological medical images, perform image denoising and image enhancement, and divide the dataset into training set and test set according to the ratio; Step 2: Constructing a Unet network based on gated aggregation and Inception. The Unet network based on gated aggregation and Inception uses an Inception-based module to extract multi-scale features and enrich lesion feature information. Then, through gated aggregation, the selectivity of the network in the feature aggregation process is enhanced, and the network is allowed to better control which feature information is aggregated. Step 3: Input the preprocessed dermatology medical image dataset into the Unet network based on gated aggregation and Inception for training to obtain the optimal model; Step 4: After the training is completed, the dermatology medical image test set is input into the obtained optimal model to detect the segmentation results of the lesions in the dermatology medical images.

2. According to claim 1, a medical image segmentation method based on gated aggregation and Inception, characterized in that: The step 1 comprises the following steps: Step 1.1: Read the color image through the image processing library and obtain its RGB channel information. The RGB image consists of three color channels: red, green, and blue, and each channel contains the corresponding pixel value; Step 1.2: The pixel values ​​of the image are normalized to the range [0,1]. The normalization formula is as follows: Step 1.3: bilateral filtering denoising; First, the spatial distance weight is calculated to take into account the geometric distance between pixels. The spatial distance weight calculation formula is as follows: Next, the weights between pixel values ​​are calculated to maintain edge information. The pixel value weight calculation formula is as follows: Then, the spatial and pixel value weights are combined to obtain the normalized weights, which are calculated as follows: Finally, the normalized weights are used to perform weighted averaging on the pixels in the neighborhood to obtain the filtered image. The formula for the filtered pixel value is as follows: Step 1.4: Image enhancement based on histogram equalization; Step 1.4.1: Convert the denoised RGB image to YCrCb color space to separate the luminance and chrominance information; Step 1.4.2: Extract the brightness channel Y from the YCrCb image and process its histogram separately to enhance the brightness contrast of the image; Step 1.4.3: Apply adaptive histogram equalization to process the brightness channel Y to enhance local contrast and improve image details; Step 1.4.4: Recombine the equalized luminance channel Y with the original chrominance channels Cr and Cb to form an enhanced YCrCb image; Step 1.4.5: Convert the processed YCrCb image back to RGB color space to obtain an enhanced image with better visual effect; Step 1.5: Divide the skin disease image dataset after data enhancement in step 1.4 into a training set and a test set in a ratio of 8:

2.

3. The medical image segmentation method based on gated aggregation and Inception according to claim 1, characterized in that: The step 2 comprises the following steps: Step 2.1: Use the improved Inception module to extract multi-scale feature information. The improved Inception module consists of four parallel branches, and each branch independently completes its own feature extraction process; Step 2.2: Branch 1; The number of channels is compressed through 1x1 convolution, and then the spatial features are extracted through 3x3 convolution. The detailed features in the horizontal and vertical directions are extracted through 3x1 and 1x3 convolution respectively. Finally, the depth-wise separable convolution is used to further reduce the amount of calculation and enhance the feature expression capability. X 1，1 =Conv 1×1 (X) X 1,2 =Conv 3×3 (X 1，1 ) X 1,3 =Conv 3×1 (X 1,2 )+Conv 1×3 (X 1,2 ) X 1,4 =DepthwiseSeparableConv(X 1,3 ) Step 2.3: Branch 2; Use 1x1 convolution to reduce the number of channels, then use 3x1 and 1x3 convolution to extract features in the horizontal and vertical directions respectively, and finally use the attention mechanism SE module to enhance important features and suppress irrelevant features; X 2，1 =Conv 1×1 (X) X 2,2 =Conv 3×1 (X 2，1 )+Conv 1×3 (X 2，1 ) X 2,3 =Attention(X 2,2 ) Step 2.4: Branch 3; First, the global context information is captured through pooling, i.e., average pooling or maximum pooling. Then, the feature channels are further compressed through 1x1 convolution. Finally, the nonlinear ability of feature expression is enhanced through bilinear pooling. X 3，1 =Pool(X) X 3,2 =Conv 1×1 (X 3，1 ) X 3,3 =BilinearPooling(X 3,2 ) Step 2.5: Branch 4; The original feature information is preserved through 1x1 convolution, and the receptive field is introduced through extended convolution to capture a wide range of contextual information; X 4，1 =Conv 1×1 (X) X 4，2 =DilatedConv(X 4，1 ) Step 2.6: Branch splicing; The outputs of branches 1, 2, 3, and 4 are concatenated in the channel dimension to form a comprehensive feature map, and then the concatenated feature map is subjected to batch normalization and ReLU activation function to normalize and activate the features; X concat =Concat(X 1,4 ,X 2,3 ,X 3,3 ,X 4,2 ) X bn =BatchNorm(X concat ) X final =ReLU(X bn ) Step 2.7: Use the gated aggregation mechanism to perform weighted fusion on the concatenated feature maps obtained in step 2.

6. First, use global average pooling to extract global information, then use the fully connected layer to generate weight coefficients, and then use the Sigmoid activation function to ensure that the weight is in the range of [0,1]. Finally, perform weighted fusion and adjust the weight of each branch feature by multiplying each channel. G=GAP(X final ) G reduced =ReLU(Conv 1×1 (G)) G restored =σ(Conv 1×1 (G reduced )) X gated =X final ⊙G restored Y final =ReLU(X gated )。 4. The medical image segmentation method based on gated aggregation and Inception according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1: Initialize the parameters of the Unet network based on gated aggregation and Inception; Step 3.2: Set the training parameters, the training epoch is 120, the single training data batch size is 4, the initial learning rate lr of the network is set to 0.0001, and the Adam optimizer is used; Step 3.3: Training uses DSC, mIoU and Accuracy as evaluation indicators of the model to evaluate the training results. The function expression of the DSC coefficient is as follows: Among them, TP, FP, and FN represent true positives, false positives, and false negatives, respectively; The function expression of the mIoU coefficient is as follows: Among them, TP, FP, and FN represent true positives, false positives, and false negatives, respectively; The function expression of the Accuracy coefficient is as follows: Among them, TP, FP, FN, and TN represent true positives, false positives, false negatives, and true negatives, respectively; Step 3.4: Use the improved UNet network model for iterative training to extract the characteristic information of skin disease images and obtain the optimal model for segmentation result output.

Images