Insulin resistance evaluation method based on multi-scale liver pathological image feature fusion

Abstract

The invention relates to an insulin resistance evaluation method based on multi-scale liver pathological image feature fusion, and belongs to the field of medical artificial intelligence. The method comprises the following steps: acquiring a liver pathological image of a TBT-induced insulin resistance animal model, and constructing a data set; processing the training set by using a CLAHE algorithm to obtain preprocessed liver data; constructing an insulin resistance grading evaluation model, wherein the model comprises a sliding window encoder, a multi-scale attention perception module, a first pyramid encoder, a second pyramid encoder, a cross-layer connection fusion module, a metabolic characteristic mapping module and a grading classifier; training the model by adopting the preprocessed liver data; performing training optimization on the model by using an Adam optimizer through a loss function to obtain a trained model; and preprocessing to-be-graded images in a test set, and inputting the to-be-graded images into the trained model to obtain a grading result. According to the invention, the accuracy and generalization ability of model grading can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical artificial intelligence, and specifically relates to an insulin resistance assessment method based on multi-scale liver pathology image feature fusion. Background Art

[0002] With economic development, changes in eating habits and lifestyles, the global obesity problem is becoming increasingly serious. Obesity not only leads to chronic low-grade inflammation, but may also cause insulin resistance and type 2 diabetes. Tributyltin (TBT), as a typical environmental obesity factor, can cause metabolic syndromes such as obesity and insulin resistance at low concentrations, but its specific mechanism of action has not yet been clarified. Previous studies have shown that TBT treatment can significantly reduce the level of insulin-induced AKT phosphorylation in animal livers. It was further found that TBT significantly reduced the expression of insulin receptor substrate (IRS) 2 upstream of the PI3K-AKT pathway, and this change was closely related to changes in macrophage polarization in the liver, suggesting that microenvironmental changes play an important role in TBT-induced insulin resistance. Existing methods for assessing insulin resistance mainly rely on biochemical indicators and cannot fully reflect the subtle changes in liver pathology. Therefore, feature analysis based on liver pathology images is of great significance.

[0003] Traditional image processing methods often only extract features at a single scale, potentially missing subtle manifestations of certain lesions. Multiscale image feature fusion technology can extract and fuse image information at multiple scales, comprehensively capturing potential lesion characteristics in liver pathology images. Multiscale analysis not only effectively improves image resolution but also enhances the recognizability of different pathological features within the image, improving the accuracy of insulin resistance grading assessments.

[0004] Furthermore, multi-scale feature fusion technology can combine image information extracted at different scales and further analyze, process, and classify it through advanced computational models. This method can extract multi-dimensional features from liver pathology images, such as hepatocyte size, morphological changes, degree of liver tissue fibrosis, and distribution of fatty infiltration. These features can not only help doctors better identify early signs of insulin resistance but also provide more accurate grading between different pathological changes. Algorithms based on multi-scale liver pathology image feature fusion not only enable automated diagnosis but also reduce errors in human judgment, providing a more objective and accurate tool for assessing insulin resistance and promoting the development of personalized and precision medicine. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides an insulin resistance assessment method based on multi-scale liver pathology image feature fusion.

[0006] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions: The present invention provides a method for evaluating insulin resistance based on multi-scale liver pathology image feature fusion, comprising the following steps: S1. Obtain liver pathological images from an animal model of TBT-induced insulin resistance, construct the LiverTBT dataset, and divide it into training and test sets. S2. Use the CLAHE algorithm to preprocess the liver images in the training set of the LiverTBT dataset to obtain preprocessed liver data; S3. Construct a hierarchical assessment model for insulin resistance, comprising a sliding window encoder, a multiscale perceptual attention module, a first pyramid encoder, a second pyramid encoder, a cross-layer connection fusion module, a metabolic feature mapping module, and a hierarchical classifier; train the model using preprocessed liver data; S4. Use the Adam optimizer to train and optimize the model through the loss function to obtain a trained model; S5. Preprocess the images to be graded in the test set and input them into the trained model to obtain the graded results.

[0007] Furthermore, in step S1, the insulin resistance level of the liver pathology image is annotated and divided into four levels from strong to weak, including: severe insulin resistance A, moderate insulin resistance B, mild insulin resistance C, and no / very mild insulin resistance D. 80% of the data is used as a training set and 20% as a test set.

[0008] Furthermore, in step S2, the parameters of the CLAHE algorithm are set according to the characteristics of the image. The tile size TileSize is set to 8x8, which means that the original liver pathology image is divided into 8 rows and 8 columns, a total of 64 small areas; the clipping threshold or contrast limiting parameter Clip Limit is set to 2.0; the interpolation process adopts the bilinear interpolation method; all images in the training set are processed by the CLAHE algorithm to obtain the preprocessed liver data .

[0009] Furthermore, the sliding window encoder in step S3 is composed of a sliding window module, a convolution layer, and a pooling layer in sequence; the window size in the sliding window module is selected to be 32x32 pixels, the stride is selected to be 16 pixels, and the overlap area between windows is set to 50%; the convolution kernel size of the convolution layer is 3×3, the number of convolution kernels is set to 32, the ReLU activation function is used to increase the nonlinear expression ability, and the stride is set to 1; the pooling layer is maximum pooling, the pooling window size is set to 2×2, and the stride is set to 2; The preprocessed liver data After processing by the sliding window encoder, the sliding coding feature is obtained .

[0010] Furthermore, the multi-scale perception attention module in step S3 includes a spatial information processing unit, a channel information processing unit, a scale information processing unit and a multi-scale attention unit; the sliding coding feature Input into the spatial information processing unit, channel information processing unit, and scale information processing unit for processing respectively to obtain spatial information , channel information , scale information ; The formula for the processing process in the spatial information processing unit is as follows: , in, Represents sliding coding features The mean of Represents sliding coding features The standard deviation of Indicates the convolution layer operation, the convolution kernel size is 3×3, and the number of convolution kernels is set to 64; Represents the RELU activation function; Represents the Sigmoid activation function; The formula for the processing process in the channel information processing unit is as follows: , in, represents a constant, ; represents element-wise multiplication, Indicates the average operation in the spatial dimension, Indicates operations along the 1st and 2nd dimensions; The formula for the processing process in the scale information processing unit is as follows: , in, represents the number of multi-scale convolution kernels, ; Indicates that different sizes of convolution kernels are used for processing. Indicates that the convolution kernel size is 3×3, Indicates that the convolution kernel size is 5×5; The spatial information , channel information , scale information Input to the multi-scale attention unit for processing to obtain multi-scale attention , the formula is as follows: , in, 、 、 represent the first trainable parameter, the second trainable parameter, and the third trainable parameter respectively; 、 、 Obey the standard normal distribution; Represents the dot product operation.

[0011] Furthermore, in step S3, the first pyramid encoder and the second pyramid encoder specifically include: The first pyramid encoder includes a first convolutional layer, a second convolutional layer, a local nonlinear interaction module, a LeakyReLU nonlinear activation function, and a downsampling module; the convolution kernel size of the first convolutional layer is 3×3, the number of convolution kernels is 128, the stride is 1, and "same" padding is used; the convolution kernel size of the second convolutional layer is 3×3, the number of convolution kernels is 256, the stride is 1, and "same" padding is used; the local nonlinear interaction module uses depthwise separable convolution, the convolution kernel size is 3×3×1, the stride is 2, and the padding mode is "VALID"; the downsampling module uses a deconvolution method, the convolution kernel size is 3×3, the stride is set to 2, the padding size is 1, and the dilation coefficient is 2; a learnable bias term is added after the convolution operation; The preprocessed liver data After being processed by the first convolution layer, the second convolution layer, the local nonlinear interaction module, and the LeakyReLU nonlinear activation function, the spatial attention Add the input to the downsampling module in the spatial dimension to obtain the first encoding feature ; The second pyramid encoder is composed of a third convolutional layer, a ReLU activation function, a fourth convolutional layer, a ReLU activation function, and a pooling layer in sequence; the convolution kernel size of the third convolutional layer is 3×3, the number of convolution kernels is 512, the stride is 1, and "same" padding is used; the convolution kernel size of the fourth convolutional layer is 5×5, the number of convolution kernels is 1024, the stride is 2, and "same" padding is used; the pooling layer adopts maximum pooling, the pooling window size is 3×3, and the stride is 1; The preprocessed liver data After processing by the second pyramid encoder, the second encoding feature is obtained .

[0012] Furthermore, the cross-layer connection fusion module in step S3 specifically includes: The cross-layer connection fusion module is composed of the fifth convolutional layer, ReLU activation function, average pooling layer, sixth convolutional layer, seventh convolutional layer, ReLU activation function, and maximum pooling layer in sequence; the convolution kernel size of the fifth convolutional layer is 1×1, the number of convolution kernels is 128, the stride is 1, and "same" padding is used; the average pooling layer adopts average pooling, the pooling window size is 1×1, and the stride is 1; the convolution kernel size of the sixth convolutional layer is 3×3, the number of convolution kernels is 256, the stride is 1, and "same" padding is used; the convolution kernel size of the seventh convolutional layer is 3×3, the number of convolution kernels is 512, the stride is 3, and "same" padding is used; the maximum pooling layer adopts maximum pooling, the pooling window size is 2×2, and the stride is 1; The second coding feature Perform pruning operation, use L1 norm as the feature importance evaluation indicator in pruning operation, and calculate the second encoding feature The L1 norm of the i-th feature map in , the formula is as follows: , in, and Represent the height and width of the feature map respectively, Represents the i-th channel in the feature map The pixel value of the position; the second encoding feature The L1 norm of the feature map is sorted from large to small, and a part of the channels are pruned from small to large, and the pruning ratio is set to 30% to obtain the pruned features. ; The first encoding feature and the second coded feature Perform element-by-element addition to obtain the third encoding feature ; The third coding feature It is sequentially input into the fifth convolutional layer, ReLU activation function, average pooling layer and sixth convolutional layer for processing to obtain the fourth encoding feature. ; Prune features and the fourth coding feature After element-by-element addition, the data is input into the seventh convolution layer, ReLU activation function, and maximum pooling layer to obtain the fusion feature. .

[0013] Furthermore, the metabolic signature mapping module includes a mapping calculation layer and a high-order convolution weighted filter; Fusion Features After normalization processing in the mapping calculation layer, the standardized features are obtained , the formula is as follows: , in, and Represents fusion features The mean and standard deviation of the standardized features Perform weighted adaptive normalization operation to obtain normalized features , the formula is as follows: , in, represents the learnable weight matrix, Represents the bias term; the normalized feature After processing with high-order convolution weighted filters, metabolic characteristics are obtained , the formula is as follows: , in, (i,j) is the convolution kernel at position The weight of Represents normalized features The pixel value at spatial location (i,j).

[0014] Furthermore, the hierarchical classifier in step S3 is composed of a flattening layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a max function in sequence; the number of units in the first fully connected layer is 512, the number of units in the second fully connected layer is 64, and the number of units in the third fully connected layer is 4; the max function increments the subscript corresponding to the maximum value in the output of the third fully connected layer by one, and then matches the result with the classification table dictionary and outputs the result; The metabolic profile and the third coded feature After the flattening layer performs element-by-element addition, it is flattened into a one-dimensional vector along the channel direction to obtain the flattened feature ; The flattening feature After the first fully connected layer, the second fully connected layer, the third fully connected layer and the max function, the insulin resistance grade is obtained .

[0015] Furthermore, the loss function in step S4 adopts the weighted cross entropy loss function, which is expressed as follows: , in, represents the weighted cross entropy loss function, represents the actual label, Represents the category probability predicted by the model; represents the number of categories, ; Representation category The weight of .

[0016] The advantages of the present invention are: This paper proposes an insulin resistance assessment method based on multi-scale liver pathology image feature fusion. It combines sliding window encoding, multi-scale attention, pyramid encoding, and channel pruning modules to achieve efficient extraction and optimization of multi-level and multi-scale features of pathology images. By focusing on key areas with spatial attention, removing redundant information through channel pruning, and adopting cross-layer feature fusion and metabolic feature mapping, the accuracy and generalization ability of the model classification are effectively improved. At the same time, this method is end-to-end automated, computationally efficient, and has strong interpretability of results. It can adapt to complex and diverse liver pathological changes and provides a scientific, accurate, and practical intelligent auxiliary tool for clinical insulin resistance classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] 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.

[0018] Figure 1 is a flow chart of the steps of the method of the present invention; Figure 2 This is the insulin resistance grade classification of liver pathology images in the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments derived by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] Example 1 In this embodiment, Figure 1 As shown, the present invention provides an insulin resistance assessment method based on multi-scale liver pathology image feature fusion, the specific steps of which include: S1. Obtain liver pathological images from an animal model of TBT-induced insulin resistance, construct the LiverTBT dataset, and divide it into training and test sets. Specifically, S11. Select mice or rats as experimental subjects to ensure the consistency of their genetic background. Ensure the consistency of the experimental group and the control group through appropriate animal management and environmental control. Determine the intraperitoneal injection of TBT treatment method based on previous studies, set a treatment time of 2 weeks, and observe the induction effect of TBT on animal insulin resistance. Verify the insulin resistance characteristics of the animal model through routine biochemical indicators such as blood glucose, insulin, and HOMA-IR index. After the experiment, anesthetize the animals and dissect out the liver. Select the cross-section cutting method according to the experimental design to ensure the representativeness of the sample. The removed liver tissue is immediately immersed in 10% formalin solution for fixation for 48 hours. The fixed liver tissue is dehydrated, transparentized and paraffin-embedded to ensure the integrity of the tissue structure. Use a microtome to cut the paraffin-embedded liver tissue into 4-5 micron thick slices, and place the slices on a slide. Select the HE staining method to stain the liver sections to highlight the structure of the liver tissue, the lesion area, and the pathological characteristics such as the inflammatory response.

[0021] S12. Use a fluorescence microscope to photograph the stained liver sections. Ensure that the images captured are clear and unblurred. Invite pathology experts to annotate the collected liver pathology images to identify different areas of liver tissue and pathological features related to insulin resistance. Annotate the insulin resistance level of liver pathology images and divide them into four levels from strong to weak, including: severe insulin resistance A, moderate insulin resistance B, mild insulin resistance C, and no / very mild insulin resistance D. Use 80% of the data as a training set and 20% as a test set.

[0022] In this method, insulin resistance is divided into four levels from strong to weak: A, B, C, and D. Figure 2 The specific meanings are as follows: Grade A (severe insulin resistance): There is obvious fatty degeneration in the liver tissue (large or small droplets of fat deposition), significant inflammatory cell infiltration, obvious ballooning of hepatocytes, a high degree of fibrosis, and pseudolobule formation can be seen in some parts.

[0023] Grade B (moderate insulin resistance): Fatty degeneration and inflammatory cell infiltration are obvious, but fibrosis is in the early stages, and ballooning and necrotic areas are less common than in Grade A.

[0024] Grade C (mild insulin resistance): Mild fatty degeneration is visible, the inflammatory reaction is mild, there is no obvious fibrosis, and the liver cell structure is basically intact.

[0025] Grade D (no / very mild insulin resistance): There is basically no fatty degeneration in the liver section, no inflammatory cell infiltration, and normal hepatocyte morphology, indicating a healthy or near-healthy state.

[0026] S2. Use the CLAHE algorithm to preprocess the liver images in the training set of the LiverTBT dataset to obtain preprocessed liver data; Specifically, the parameters of the CLAHE algorithm are set based on the characteristics of the image. Tile Size refers to the division of the entire image into several small subregions (i.e., tiles) when applying the CLAHE algorithm, with each subregion undergoing a separate histogram equalization operation. In this method, the tile size is set to 8x8, indicating that the original liver pathology image is divided into 8 rows and 8 columns, totaling 64 small regions. This enhances local image contrast, making detailed features more prominent and facilitating subsequent analysis. The Clip Limit parameter in the CLAHE algorithm is used to limit the maximum number of pixels at a given grayscale level within the histogram of each subregion. If the number of pixels at a given grayscale level exceeds this threshold, they are "clipped," and the excess is evenly distributed across other grayscale levels. This effectively prevents noise amplification or over-enhancement caused by overly concentrated local pixel distribution. In this method, the clipping threshold, or contrast limiting parameter, Clip Limit, is set to 2.0, which helps control the equalization effect, improve image quality, and avoid artifacts. Bilinear interpolation is used in the interpolation process; all images in the training set are processed by the CLAHE algorithm to obtain preprocessed liver data. .

[0027] S3. Construct a hierarchical assessment model for insulin resistance, comprising a sliding window encoder, a multiscale perceptual attention module, a first pyramid encoder, a second pyramid encoder, a cross-layer connection fusion module, a metabolic feature mapping module, and a hierarchical classifier; train the model using preprocessed liver data; Specifically, the sliding window encoder is composed of a sliding window module, a convolution layer, and a pooling layer in sequence; the window size in the sliding window module is selected as 32x32 pixels, the stride is selected as 16 pixels, and the overlapping area between windows is set to 50%; the convolution kernel size of the convolution layer is 3×3, the number of convolution kernels is set to 32, the ReLU activation function is used to increase the nonlinear expression ability, and the stride is set to 1; the pooling layer is maximum pooling, the pooling window size is set to 2×2, and the stride is set to 2; the preprocessed liver data After processing by the sliding window encoder, the sliding coding feature is obtained .

[0028] Specifically, the multi-scale perception attention module includes a spatial information processing unit, a channel information processing unit, a scale information processing unit and a multi-scale attention unit; the sliding coding feature Input into the spatial information processing unit, channel information processing unit, and scale information processing unit for processing respectively to obtain spatial information , channel information , scale information ; The formula for the processing process in the spatial information processing unit is as follows: , in, Represents sliding coding features The mean of Represents sliding coding features The standard deviation of Indicates the convolution layer operation, the convolution kernel size is 3×3, and the number of convolution kernels is set to 64; Represents the RELU activation function; Represents the Sigmoid activation function; The formula for the processing process in the channel information processing unit is as follows: , in, represents a constant, ; Used to avoid numerical problems when calculating logarithms, represents element-wise multiplication, Indicates the average operation in the spatial dimension, Indicates operations along the 1st and 2nd dimensions; The formula for the processing process in the scale information processing unit is as follows: , in, represents the number of multi-scale convolution kernels, ; Indicates that different sizes of convolution kernels are used for processing. Indicates that the convolution kernel size is 3×3, Indicates that the convolution kernel size is 5×5; The spatial information , channel information , scale information Input to the multi-scale attention unit for processing to obtain multi-scale attention , the formula is as follows: , in, 、 、 represent the first trainable parameter, the second trainable parameter, and the third trainable parameter respectively; 、 、 Obey the standard normal distribution; Represents the dot product operation.

[0029] Specifically, the first pyramid encoder includes a first convolutional layer, a second convolutional layer, a local nonlinear interaction module, a LeakyReLU nonlinear activation function, and a downsampling module; the convolution kernel size of the first convolutional layer is 3×3, the number of convolution kernels is 128, the stride is 1, and "same" padding is used; the convolution kernel size of the second convolutional layer is 3×3, the number of convolution kernels is 256, the stride is 1, and "same" padding is used; the local nonlinear interaction module uses depthwise separable convolution, the convolution kernel size is 3×3×1, the stride is 2, and the padding mode is "VALID"; the downsampling module uses the deconvolution method ConvTranspose(), the convolution The kernel size kernel_size is 3×3, and the stride is set to 2, that is, 2 pixels are moved on the input feature map during each convolution operation, which makes the size of the output feature map approximately twice that of the input, achieving an upsampling effect; the padding size padding is 1, that is, a circle of zeros is padded on the edge of the input feature map, so that the output feature map retains the boundary information while the size is expanded; the dilation coefficient dilation is 2, that is, the interval between adjacent elements in the convolution kernel is 2 pixels, which can expand the receptive field and obtain richer contextual information; adding a learnable bias term bias=1 after the convolution operation is beneficial to improving the model's expressiveness and convergence speed.

[0030] The preprocessed liver data After being processed by the first convolution layer, the second convolution layer, the local nonlinear interaction module, and the LeakyReLU nonlinear activation function, the spatial attention Add the input to the downsampling module in the spatial dimension to obtain the first encoding feature ; The second pyramid encoder is composed of a third convolutional layer, a ReLU activation function, a fourth convolutional layer, a ReLU activation function, and a pooling layer in sequence; the convolution kernel size of the third convolutional layer is 3×3, the number of convolution kernels is 512, the stride is 1, and "same" padding is used; the convolution kernel size of the fourth convolutional layer is 5×5, the number of convolution kernels is 1024, the stride is 2, and "same" padding is used; the pooling layer adopts maximum pooling, the pooling window size is 3×3, and the stride is 1; The preprocessed liver data After processing by the second pyramid encoder, the second encoding feature is obtained .

[0031] Specifically, the cross-layer connection fusion module is composed of a fifth convolutional layer, a ReLU activation function, an average pooling layer, a sixth convolutional layer, a seventh convolutional layer, a ReLU activation function, and a maximum pooling layer in sequence; the convolution kernel size of the fifth convolutional layer is 1×1, the number of convolution kernels is 128, the stride is 1, and "same" is used for padding; the average pooling layer adopts average pooling, the pooling window size is 1×1, and the stride is 1; the convolution kernel size of the sixth convolutional layer is 3×3, the number of convolution kernels is 256, the stride is 1, and "same" is used for padding; the convolution kernel size of the seventh convolutional layer is 3×3, the number of convolution kernels is 512, the stride is 3, and "same" is used for padding; the maximum pooling layer adopts maximum pooling, the pooling window size is 2×2, and the stride is 1; The second coding feature A pruning operation is performed to completely remove some of the channels with the lowest ranking according to the importance index of the L1 norm, and only retain the most important feature channels, so as to achieve the purpose of streamlining the model, speeding up the inference speed and reducing the storage requirement.

[0032] In the pruning operation, the L1 norm is used as the feature importance evaluation indicator to calculate the second encoding feature The L1 norm of the i-th feature map in , the formula is as follows: , in, and Represent the height and width of the feature map respectively, Represents the i-th channel in the feature map The pixel value of the position; the second encoding feature The L1 norm of the feature maps in the image are sorted from large to small, and a part of the channels are pruned from small to large. The pruning ratio is set to 30%. The channels corresponding to the 30% feature maps with the smallest L1 norm after sorting are removed from the entire feature tensor, leaving only the remaining 70%. The removed channels no longer participate in the forward propagation and parameter update of the subsequent network. ; The first encoding feature and the second coded feature Perform element-by-element addition to obtain the third encoding feature ; The third coding feature It is sequentially input into the fifth convolutional layer, ReLU activation function, average pooling layer and sixth convolutional layer for processing to obtain the fourth encoding feature. ; Prune features and the fourth coding feature After element-by-element addition, the data is input into the seventh convolution layer, ReLU activation function, and maximum pooling layer to obtain the fusion feature. .

[0033] Specifically, the metabolic signature mapping module includes a mapping calculation layer and a high-order convolution weighted filter; Fusion Features After normalization processing in the mapping calculation layer, the standardized features are obtained , the formula is as follows: , in, and Represents fusion features The mean and standard deviation of the standardized features Perform weighted adaptive normalization operation to obtain normalized features , the formula is as follows: , in, represents the learnable weight matrix, Represents the bias term; the normalized feature After processing with high-order convolution weighted filters, metabolic characteristics are obtained , the formula is as follows: , in, (i,j) is the convolution kernel at position The weight of Represents normalized features The pixel value at spatial location (i,j).

[0034] Specifically, the hierarchical classifier is composed of a flattening layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a max function in sequence; the number of units in the first fully connected layer is 512, the number of units in the second fully connected layer is 64, and the number of units in the third fully connected layer is 4; the max function increments the subscript corresponding to the maximum value in the output of the third fully connected layer by one, and then matches the output with the classification table dictionary; The metabolic profile and the third coded feature After adding through the flattening layer, it is flattened into a one-dimensional vector along the channel direction to obtain the flattened feature ; The flattening feature After the first fully connected layer, the second fully connected layer, the third fully connected layer and the max function, the insulin resistance grade is obtained .

[0035] Assume that the output of the third fully connected layer is [0.1, 0.3, 2.7, 0.9], then: The maximum value is 2.7, corresponding to the subscript 2; Add 1 to subscript 2 to get category number 3; If the classification table dictionary is defined as {1: "Class A", 2: "Class B", 3: "Class C", 4: "Class D"}, the output is "Class C".

[0036] S4. Use the Adam optimizer to train and optimize the model through the loss function to obtain a trained model; Specifically, the loss function adopts the weighted cross entropy loss function, which is expressed as follows: , in, represents the weighted cross entropy loss function, represents the actual label, represents the category probability predicted by the model, is the number of categories, is a category The weight of is inversely proportional to the sample frequency of the category and is the inverse of the sample frequency of the category to balance the class imbalance.

[0037] S5. Preprocess the images to be graded in the test set and input them into the trained model to obtain the graded results.

[0038] Example 2 In this embodiment, a comparative experiment was conducted between the method of the present invention and several existing advanced classification methods to verify the effectiveness and advantages of the method of the present invention. The LiverTBT dataset, containing liver pathology images from an animal model of TBT-induced insulin resistance, was used. Existing methods include: CNN (convolutional neural network), widely used for image classification; SVM (support vector machine), commonly used for small-sample learning and high-dimensional data classification; KNN (k-nearest neighbor algorithm), a simple and easy-to-understand instance-based learning method; and BPNN (back-propagation neural network), a classic multi-layer feedforward neural network. Evaluation metrics included accuracy, recall, precision, F1 score, and AUC. Each method was trained and tested on the LiverTBT dataset to assess its performance in the insulin resistance grading task. Experimental results show that this method outperforms other methods in terms of accuracy, recall, precision, F1 score, and AUC (area under the receiver operating characteristic (ROC) curve). This method particularly demonstrates outstanding performance in terms of AUC and F1 score, demonstrating its high accuracy and robustness in the insulin resistance grading task. Compared with CNN, this method comprehensively considers the characteristics of insulin resistance, not just low-level image features, resulting in improved performance. Compared with SVM, this method automatically learns features without the need for manual kernel selection and performs better when processing complex data. Compared with KNN, this method is more advantageous when processing high-dimensional data and is not affected by distance metrics. Compared with BPNN, this method better captures detailed image information through multi-scale feature fusion and metabolic feature mapping. These comparative experiments demonstrate that this method performs well in the insulin resistance grading task, outperforming existing mainstream classification methods. Its multi-scale feature fusion and metabolic feature mapping design facilitate more accurate identification of different levels of insulin resistance, providing strong support for clinical diagnosis.

[0039] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for evaluating insulin resistance based on multi-scale liver pathology image feature fusion, characterized in that: The following steps are involved: S1. Obtain liver pathological images from an animal model of TBT-induced insulin resistance, construct the LiverTBT dataset, and divide it into training and test sets. S2. Use the CLAHE algorithm to preprocess the liver images in the training set of the LiverTBT dataset to obtain preprocessed liver data; S3. Construct a hierarchical assessment model for insulin resistance, comprising a sliding window encoder, a multiscale perceptual attention module, a first pyramid encoder, a second pyramid encoder, a cross-layer connection fusion module, a metabolic feature mapping module, and a hierarchical classifier; train the model using preprocessed liver data; S4. Use the Adam optimizer to train and optimize the model through the loss function to obtain a trained model; S5. Preprocess the images to be graded in the test set and input them into the trained model to obtain the graded results.

2. The insulin resistance assessment method based on multi-scale liver pathology image feature fusion according to claim 1, characterized in that: In step S1, the insulin resistance level of the liver pathology image is annotated and divided into four levels from strong to weak, including: severe insulin resistance A, moderate insulin resistance B, mild insulin resistance C, and no / very mild insulin resistance D. 80% of the data is used as a training set and 20% as a test set.

3. The insulin resistance assessment method based on multi-scale liver pathology image feature fusion according to claim 2, characterized in that: In step S2, the parameters of the CLAHE algorithm are set according to the characteristics of the image. The tile size is set to 8x8, which means that the original liver pathology image is divided into 8 rows and 8 columns, a total of 64 small areas; the clipping threshold or contrast limiting parameter Clip Limit is set to 2.0; the interpolation process uses the bilinear interpolation method; all images in the training set are processed by the CLAHE algorithm to obtain the preprocessed liver data .

4. The insulin resistance assessment method based on multi-scale liver pathology image feature fusion according to claim 3, characterized in that: The sliding window encoder in step S3 specifically includes: The sliding window encoder is composed of a sliding window module, a convolutional layer, and a pooling layer in sequence; the window size of the sliding window module is selected to be 32×32 pixels, the stride is selected to be 16 pixels, and the overlap area between windows is set to 50%; the convolution kernel size of the convolution layer is 3×3, the number of convolution kernels is set to 32, the ReLU activation function is used to increase the nonlinear expression ability, and the stride is set to 1; the pooling layer is a maximum pooling, the pooling window size is set to 2×2, and the stride is set to 2; The preprocessed liver data After processing by the sliding window encoder, the sliding coding feature is obtained .

5. The insulin resistance assessment method based on multi-scale liver pathology image feature fusion according to claim 4, characterized in that: The multi-scale perception attention module in step S3 specifically includes: The multi-scale perception attention module includes a spatial information processing unit, a channel information processing unit, a scale information processing unit and a multi-scale attention unit; the sliding coding feature Input into the spatial information processing unit, channel information processing unit, and scale information processing unit for processing respectively to obtain spatial information , channel information , scale information ; The formula for the processing process in the spatial information processing unit is as follows: ， in, Represents sliding coding features The mean of Represents sliding coding features The standard deviation of Indicates the convolution layer operation, the convolution kernel size is 3×3, and the number of convolution kernels is set to 64; Represents the RELU activation function; Represents the Sigmoid activation function; The formula for the processing process in the channel information processing unit is as follows: ， in, represents a constant, ; represents element-wise multiplication, Indicates the average operation in the spatial dimension, Indicates operations along the 1st and 2nd dimensions; The formula for the processing process in the scale information processing unit is as follows: ， in, represents the number of multi-scale convolution kernels, ; Indicates that different sizes of convolution kernels are used for processing. Indicates that the convolution kernel size is 3×3, Indicates that the convolution kernel size is 5×5; The spatial information , channel information , scale information Input to the multi-scale attention unit for processing to obtain multi-scale attention , the formula is as follows: ， in, 、 、 represent the first trainable parameter, the second trainable parameter, and the third trainable parameter respectively; 、 、 Obey the standard normal distribution; Represents the dot product operation.

6. The insulin resistance assessment method based on multi-scale liver pathology image feature fusion according to claim 5, characterized in that: In step S3, the first pyramid encoder and the second pyramid encoder specifically include: The first pyramid encoder includes a first convolutional layer, a second convolutional layer, a local nonlinear interaction module, a LeakyReLU nonlinear activation function, and a downsampling module; the convolution kernel size of the first convolutional layer is 3×3, the number of convolution kernels is 128, the stride is 1, and "same" padding is used; the convolution kernel size of the second convolutional layer is 3×3, the number of convolution kernels is 256, the stride is 1, and "same" padding is used; the local nonlinear interaction module uses depthwise separable convolution, the convolution kernel size is 3×3×1, the stride is 2, and the padding mode is "VALID"; the downsampling module uses a deconvolution method, the convolution kernel size is 3×3, the stride is set to 2, the padding size is 1, and the dilation coefficient is 2; a learnable bias term is added after the convolution operation; The preprocessed liver data After being processed by the first convolution layer, the second convolution layer, the local nonlinear interaction module, and the LeakyReLU nonlinear activation function, the spatial attention Add the input to the downsampling module in the spatial dimension to obtain the first encoding feature ; The second pyramid encoder is composed of a third convolutional layer, a ReLU activation function, a fourth convolutional layer, a ReLU activation function, and a pooling layer in sequence; the convolution kernel size of the third convolutional layer is 3×3, the number of convolution kernels is 512, the stride is 1, and "same" padding is used; the convolution kernel size of the fourth convolutional layer is 5×5, the number of convolution kernels is 1024, the stride is 2, and "same" padding is used; the pooling layer adopts maximum pooling, the pooling window size is 3×3, and the stride is 1; The preprocessed liver data After processing by the second pyramid encoder, the second encoding feature is obtained .

7. The insulin resistance assessment method based on multi-scale liver pathology image feature fusion according to claim 6, characterized in that: The cross-layer connection fusion module in step S3 specifically includes: The cross-layer connection fusion module is composed of a fifth convolutional layer, a ReLU activation function, an average pooling layer, a sixth convolutional layer, a seventh convolutional layer, a ReLU activation function, and a maximum pooling layer in sequence; the convolution kernel size of the fifth convolutional layer is 1×1, the number of convolution kernels is 128, the stride is 1, and "same" padding is used; the average pooling layer adopts average pooling, the pooling window size is 1×1, and the stride is 1; the convolution kernel size of the sixth convolutional layer is 3×3, the number of convolution kernels is 256, the stride is 1, and "same" padding is used; the convolution kernel size of the seventh convolutional layer is 3×3, the number of convolution kernels is 512, the stride is 3, and "same" padding is used; the maximum pooling layer adopts maximum pooling, the pooling window size is 2×2, and the stride is 1; The second coding feature Perform pruning operation, use L1 norm as the feature importance evaluation indicator in pruning operation, and calculate the second encoding feature The L1 norm of the i-th feature map in , the formula is as follows: ， in, and Represent the height and width of the feature map respectively, Represents the i-th channel in the feature map The pixel value of the position; the second encoding feature The L1 norm of the feature map is sorted from large to small, and a part of the channels are pruned from small to large, and the pruning ratio is set to 30% to obtain the pruned features. ; The first encoding feature and the second coded feature Perform element-by-element addition to obtain the third encoding feature ; The third coding feature It is sequentially input into the fifth convolutional layer, ReLU activation function, average pooling layer and sixth convolutional layer for processing to obtain the fourth encoding feature. ; Prune features and the fourth coding feature After element-by-element addition, the data is input into the seventh convolution layer, ReLU activation function, and maximum pooling layer to obtain the fusion feature. .

8. The insulin resistance assessment method based on multi-scale liver pathology image feature fusion according to claim 7, characterized in that: The metabolic signature mapping module in step S3 specifically includes: The metabolic feature mapping module includes a mapping calculation layer and a high-order convolution weighted filter; Fusion Features After normalization processing in the mapping calculation layer, the standardized features are obtained , the formula is as follows: ， in, and Represents fusion features The mean and standard deviation of the standardized features Perform weighted adaptive normalization operation to obtain normalized features , the formula is as follows: ， in, represents the learnable weight matrix, Represents the bias term; the normalized feature After processing with high-order convolution weighted filters, metabolic characteristics are obtained , the formula is as follows: ， in, (i,j) is the convolution kernel at position The weight of Represents normalized features The pixel value at spatial location (i,j).

9. The method for evaluating insulin resistance based on multi-scale liver pathology image feature fusion according to claim 8, characterized in that: The hierarchical classifier in step S3 specifically includes: The hierarchical classifier is composed of a flattening layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a max function in sequence; the number of units in the first fully connected layer is 512, the number of units in the second fully connected layer is 64, and the number of units in the third fully connected layer is 4; the max function adds one to the subscript corresponding to the maximum value in the output of the third fully connected layer, matches it with the classification table dictionary, and outputs it; The metabolic profile and the third coded feature After the flattening layer performs element-by-element addition, it is flattened into a one-dimensional vector along the channel direction to obtain the flattened feature ; The flattening feature After the first fully connected layer, the second fully connected layer, the third fully connected layer and the max function, the insulin resistance grade is obtained .

10. The insulin resistance assessment method based on multi-scale liver pathology image feature fusion according to claim 9, characterized in that: The loss function in step S4 adopts the weighted cross entropy loss function, which is expressed as follows: ， in, represents the weighted cross entropy loss function, represents the actual label, Represents the category probability predicted by the model; represents the number of categories, ; Representation category The weight of .

Images