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

Abstract

The present application relates to a kind of insulin resistance evaluation method based on multi-scale liver pathological image feature fusion, belong to medical artificial intelligence field.It includes the following steps: obtaining the liver pathological image of TBT-induced insulin resistance animal model, constructs data set;Using CLAHE algorithm to process training set, obtain preprocessed liver data;Insulin resistance grading evaluation model is constructed, model includes sliding window encoder, multi-scale perception attention module, first pyramid encoder, second pyramid encoder, cross-layer connection fusion module, metabolic feature mapping module and grading classifier;Preprocessed liver data is used to train model;Through loss function, model is trained and optimized using Adam optimizer, obtain trained model;Grading image in test set is preprocessed after input into trained model, obtain grading result.The present application can effectively improve the accuracy and generalization ability of model grading.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence, specifically relating to a method for assessing insulin resistance based on the fusion of multi-scale liver pathological image features. Background Technology

[0002] With economic development and changes in dietary habits and lifestyles, the global obesity problem is becoming increasingly serious. Obesity not only leads to chronic low-grade inflammation but may also trigger insulin resistance and type 2 diabetes. Tributyltin (TBT), as a typical environmental obesity factor, can induce metabolic syndrome such as obesity and insulin resistance even at low concentrations, but its specific mechanism of action remains unclear. Previous studies have shown that TBT treatment can significantly reduce insulin-induced AKT phosphorylation levels in animal livers. Further research revealed that TBT significantly reduces the expression of insulin receptor substrate (IRS)2 upstream of the PI3K-AKT pathway, and this change is closely related to altered macrophage polarization in the liver, suggesting that changes in the microenvironment play an important role in TBT-induced insulin resistance. Existing methods for assessing insulin resistance mainly rely on biochemical indicators, which cannot comprehensively reflect subtle changes in liver pathology. Therefore, feature analysis based on liver pathological images is of great significance.

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

[0004] Furthermore, multi-scale feature fusion technology can combine image information extracted at different scales and further analyze, process, and classify it using advanced computational models. This method can extract multi-dimensional features from liver pathology images, such as hepatocyte size and morphological changes, the degree of fibrosis in liver tissue, and the distribution of fatty infiltration. These features 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 can not only achieve 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 medicine and precision medicine. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method for assessing insulin resistance based on the fusion of multi-scale liver pathological image features.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] This invention provides a method for assessing insulin resistance based on the fusion of multi-scale liver pathological image features, comprising the following steps:

[0008] S1. Obtain liver pathological images of TBT-induced insulin resistance animal models, construct the LiverTBT dataset, and divide it into training and test sets;

[0009] S2. Use the CLAHE algorithm to preprocess the liver images in the training set of the LiverTBT dataset to obtain preprocessed liver data;

[0010] S3. Construct an insulin resistance grading assessment model, which includes a sliding window encoder, a multi-scale perceptual attention module, a first pyramid encoder, a second pyramid encoder, a cross-layer connection fusion module, a metabolic feature mapping module, and a grading classifier; train the model using preprocessed liver data;

[0011] S4. Use the Adam optimizer to train and optimize the model using the loss function to obtain a well-trained model;

[0012] S5. After preprocessing the images to be classified in the test set, input them into the trained model to obtain the classification results.

[0013] Furthermore, in step S1, the insulin resistance level of the liver pathology images is labeled and divided into four levels from strong to weak: 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 the training set and 20% is used as the test set.

[0014] Further, in step S2, the parameters of the CLAHE algorithm are set according to the characteristics of the image. The TileSize is set to 8x8, indicating that the original liver pathology image is divided into 8 rows and 8 columns, totaling 64 small regions; the Clip Limit parameter is set to 2.0; the bilinear interpolation method is used for interpolation; and the preprocessed liver data is obtained after all images in the training set are processed by the CLAHE algorithm. .

[0015] Further, the sliding window encoder in step S3 consists of a sliding window module, a convolutional layer, and a pooling layer in sequence; the sliding window module has a window size of 32x32 pixels, a stride of 16 pixels, and an overlap area of ​​50% between windows; the convolutional layer has a kernel size of 3×3, a kernel number of 32, uses the ReLU activation function to increase nonlinear expressive power, and has a stride of 1; the pooling layer uses max pooling, with a pooling window size of 2×2 and a stride of 2.

[0016] The preprocessed liver data After processing by a sliding window encoder, the sliding coded features are obtained. .

[0017] 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; sliding coding features The data are input into the spatial information processing unit, the channel information processing unit, and the scale information processing unit for processing to obtain spatial information. Channel information Scale information ;

[0018] The formulas for the processing in the spatial information processing unit are as follows:

[0019] ,

[0020] in, Represents sliding coding features The mean; Represents sliding coding features Standard deviation; This indicates a convolutional layer operation with a kernel size of 3×3 and a kernel count of 64. Represents the ReLU activation function; This represents the Sigmoid activation function;

[0021] The formula for the processing procedure in the channel information processing unit is as follows:

[0022] ,

[0023] in, Represents a constant. ; Represents element-wise multiplication. This indicates an averaging operation performed across spatial dimensions. This indicates that the operation is performed along the first and second dimensions;

[0024] The formula for the processing in the scale information processing unit is as follows:

[0025] ,

[0026] in, This indicates the number of multi-scale convolution kernels. ; This indicates that different sizes of convolution kernels are used for processing. This indicates that the kernel size is 3×3. This indicates that the kernel size is 5×5;

[0027] Spatial information Channel information Scale information The input is processed by a multi-scale attention unit to obtain multi-scale attention. The formula is expressed as follows:

[0028] ,

[0029] in, , , These represent the first trainable parameter, the second trainable parameter, and the third trainable parameter, respectively. , , It follows a standard normal distribution; This indicates the dot product operation.

[0030] Furthermore, the first pyramid encoder and the second pyramid encoder in step S3 specifically include:

[0031] 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 first convolutional layer has a 3×3 kernel size, 128 kernels, a stride of 1, and uses "same" padding. The second convolutional layer has a 3×3 kernel size, 256 kernels, a stride of 1, and uses "same" padding. The local nonlinear interaction module uses depthwise separable convolution with a 3×3×1 kernel size, a stride of 2, and "VALID" padding. The downsampling module uses deconvolution with a 3×3 kernel size, a stride of 2, a padding size of 1, and a dilation factor of 2. A learnable bias term is added after the convolution operation.

[0032] The preprocessed liver data After being processed sequentially through the first convolutional layer, the second convolutional layer, the local nonlinear interaction module, and the LeakyReLU nonlinear activation function, it is then combined with spatial attention. The features are added together in the spatial dimension and then input into the downsampling module to obtain the first encoded features. ;

[0033] The second pyramid encoder consists of a third convolutional layer, a ReLU activation function, a fourth convolutional layer, a ReLU activation function, and a pooling layer. The third convolutional layer has a kernel size of 3×3, 512 kernels, a stride of 1, and uses "same" padding. The fourth convolutional layer has a kernel size of 5×5, 1024 kernels, a stride of 2, and uses "same" padding. The pooling layer uses max pooling with a pooling window size of 3×3 and a stride of 1.

[0034] The preprocessed liver data After processing by the second pyramid encoder, the second encoded feature is obtained. .

[0035] Furthermore, the cross-layer connection fusion module in step S3 specifically includes:

[0036] The cross-layer connection fusion module consists 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 max pooling layer. The fifth convolutional layer has a 1×1 kernel size, 128 kernels, a stride of 1, and uses "same" padding. The average pooling layer uses average pooling with a 1×1 pooling window size and a stride of 1. The sixth convolutional layer has a 3×3 kernel size, 256 kernels, a stride of 1, and uses "same" padding. The seventh convolutional layer has a 3×3 kernel size, 512 kernels, a stride of 3, and uses "same" padding. The max pooling layer uses max pooling with a 2×2 pooling window size and a stride of 1.

[0037] For the second coding feature Pruning is performed, and the L1 norm is used as the feature importance evaluation index to calculate the second encoded feature. L1 norm of the i-th feature map The formula is expressed as follows:

[0038] ,

[0039] in, and These represent the height and width of the feature map, respectively. Represents the i-th channel in the feature map The pixel value of the location; for the second encoded feature The L1 norms of the feature maps are sorted from largest to smallest, and a portion of the channels are pruned from smallest to largest, with a pruning ratio of 30%, to obtain the pruned features. ;

[0040] The first encoded feature Second coding features By adding elements one by one, the third encoded feature is obtained. The third encoding feature The input is sequentially fed into the fifth convolutional layer, the ReLU activation function, the average pooling layer, and the sixth convolutional layer for processing to obtain the fourth encoded feature. ; Pruning characteristics and fourth coding features After element-wise addition, the results are fed into the seventh convolutional layer, the ReLU activation function, and the max pooling layer to obtain the fused features. .

[0041] Furthermore, the metabolic feature mapping module includes a mapping calculation layer and a high-order convolutional weighted filter;

[0042] Fusion features After standardization processing by the mapping computation layer, standardized features are obtained. The formula is expressed as follows:

[0043] ,

[0044] in, and They represent the fusion features respectively. The mean and standard deviation; then the standardized features Perform weighted adaptive normalization to obtain normalized features. The formula is expressed as follows:

[0045] ,

[0046] in, This represents the learnable weight matrix. Indicates the bias term; the normalized feature Metabolic features are obtained after processing with a high-order convolutional weighted filter. The formula is expressed as follows:

[0047] ,

[0048] in, (i,j) represents the position of the convolution kernel. The weights; Represents normalized features The pixel value at spatial location (i,j).

[0049] Further, the hierarchical classifier in step S3 consists of a flattening layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a max function; the first fully connected layer has 512 units, the second fully connected layer has 64 units, and the third fully connected layer has 4 units; the max function increments the index of the maximum value in the output of the third fully connected layer by one and then matches it with the classification dictionary for output.

[0050] The metabolic characteristics and third coding features After element-wise addition through the flattening layer, the vector is flattened along the channel direction to obtain a one-dimensional vector, thus yielding the flattened feature. The flattening feature After passing through 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. .

[0051] Furthermore, in step S4, the loss function adopted is the weighted cross-entropy loss function, which is expressed by the following formula:

[0052] ,

[0053] in, This represents the weighted cross-entropy loss function. Indicates the actual label, This represents the class probability predicted by the model; Indicates the number of categories, ; Indicates category The weight.

[0054] The advantages of this invention are:

[0055] This invention proposes a method for assessing insulin resistance based on multi-scale liver pathological image feature fusion. Combining modules such as sliding window coding, multi-scale attention, pyramid coding, and channel pruning, it achieves efficient extraction and optimization of multi-level, multi-scale features from pathological images. By focusing on key regions through spatial attention, removing redundant information through channel pruning, and employing cross-layer feature fusion and metabolic feature mapping, the accuracy and generalization ability of the model's grading are effectively improved. Furthermore, this method is end-to-end automated, computationally efficient, and provides highly interpretable results, adapting to complex and diverse liver pathological changes. It offers a scientific, accurate, and practical intelligent auxiliary tool for clinical insulin resistance grading. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0057] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0058] Figure 2 This invention classifies the insulin resistance level of liver pathology images. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Example 1

[0061] In this embodiment, as Figure 1 As shown, this invention provides a method for assessing insulin resistance based on the fusion of multi-scale liver pathological image features, the specific steps of which include:

[0062] S1. Obtain liver pathological images of TBT-induced insulin resistance animal models, construct the LiverTBT dataset, and divide it into training and test sets;

[0063] Specifically, S11. Mice or rats were selected as experimental subjects to ensure genetic consistency. Appropriate animal management and environmental control ensured consistency between the experimental and control groups. Based on previous research, the intraperitoneal injection of TBT was administered, with a 2-week treatment period. The induction of insulin resistance by TBT in animals was observed. Insulin resistance characteristics of the animal model were verified using routine biochemical indicators such as blood glucose, insulin, and HOMA-IR index. After the experiment, the animals were anesthetized, and the liver was dissected and removed. A transverse sectioning method was selected according to the experimental design to ensure sample representativeness. The removed liver tissue was immediately fixed in 10% formalin solution for 48 hours. The fixed liver tissue was dehydrated, cleared, and embedded in paraffin to ensure structural integrity. The paraffin-embedded liver tissue was sliced ​​into 4-5 micrometer thick sections using a microtome and placed on glass slides. HE staining was used to stain the liver sections to highlight the structure, lesion areas, and pathological features such as inflammatory reactions.

[0064] S12. Use a fluorescence microscope to photograph the stained liver sections. Ensure the images are clear and unblurred. Invite a pathology expert to annotate the acquired liver pathology images, identifying different regions of the liver tissue and pathological features related to insulin resistance. Label the insulin resistance levels of the liver pathology images, classifying them into four levels from strongest to weakest: 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 the training set and 20% as the test set.

[0065] In this method, insulin resistance is classified into four levels, A, B, C, and D, from strongest to weakest. Figure 2 As shown, the specific meaning is as follows:

[0066] Grade A (severe insulin resistance): The liver tissue shows obvious steatosis (large or small droplet fat deposition), significant inflammatory cell infiltration, obvious ballooning degeneration of hepatocytes, high degree of fibrosis, and some pseudolobule formation.

[0067] Grade B (moderate insulin resistance): Both fatty degeneration and inflammatory cell infiltration are more obvious, but fibrosis is in the early stage, and ballooning degeneration and necrotic areas are less common than in Grade A.

[0068] Grade C (mild insulin resistance): Mild steatosis is visible, with mild inflammatory response, no obvious fibrotic changes, and the hepatocyte structure is basically intact.

[0069] Grade D (No / Very Mild Insulin Resistance): Liver sections show virtually no fatty degeneration, no inflammatory cell infiltration, and normal hepatocyte morphology, indicating a healthy or near-healthy state.

[0070] S2. Use the CLAHE algorithm to preprocess the liver images in the training set of the LiverTBT dataset to obtain preprocessed liver data;

[0071] Specifically, the parameters of the CLAHE algorithm are set according to the characteristics of the image. Tile Size refers to dividing the entire image into several small sub-regions (i.e., tiles) when applying the CLAHE algorithm, with each sub-region undergoing histogram equalization separately. In this method, the tile size is set to 8x8, meaning the original liver pathology image is divided into 8 rows and 8 columns, totaling 64 small regions. This enhances the local contrast of the image, making details more prominent and facilitating subsequent analysis. Clip Limit in the CLAHE algorithm is used to limit the maximum number of pixels at a certain gray level in the histogram of each sub-region. If the number of pixels at a certain gray level exceeds this threshold, it will be "cropped," and the excess will be evenly distributed to other gray levels. This effectively prevents noise amplification or over-enhancement caused by excessive concentration of local pixels. In this method, the clipping threshold or contrast limit parameter Clip Limit is set to 2.0, which helps control the equalization effect, improve image quality, and avoid artifacts. The interpolation process uses bilinear interpolation; all images in the training set are processed by the CLAHE algorithm to obtain preprocessed liver data. .

[0072] S3. Construct an insulin resistance grading assessment model, which includes a sliding window encoder, a multi-scale perceptual attention module, a first pyramid encoder, a second pyramid encoder, a cross-layer connection fusion module, a metabolic feature mapping module, and a grading classifier; train the model using preprocessed liver data;

[0073] Specifically, the sliding window encoder consists of a sliding window module, a convolutional layer, and a pooling layer. The sliding window module has a window size of 32x32 pixels, a stride of 16 pixels, and an overlap of 50% between windows. The convolutional layer has a 3×3 kernel size, 32 kernels, uses the ReLU activation function to increase non-linear expression, and has a stride of 1. The pooling layer uses max pooling, with a pooling window size of 2×2 and a stride of 2. The preprocessed liver data... After processing by a sliding window encoder, the sliding coded features are obtained. .

[0074] 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; sliding coding features. The data are input into the spatial information processing unit, the channel information processing unit, and the scale information processing unit for processing to obtain spatial information. Channel information Scale information ;

[0075] The formulas for the processing in the spatial information processing unit are as follows:

[0076] ,

[0077] in, Represents sliding coding features The mean; Represents sliding coding features Standard deviation; This indicates a convolutional layer operation with a kernel size of 3×3 and a kernel count of 64. Represents the ReLU activation function; This represents the Sigmoid activation function;

[0078] The formula for the processing procedure in the channel information processing unit is as follows:

[0079] ,

[0080] in, Represents a constant. Used to avoid numerical problems in logarithmic calculations. Represents element-wise multiplication. This indicates an averaging operation performed across spatial dimensions. This indicates that the operation is performed along the first and second dimensions;

[0081] The formula for the processing in the scale information processing unit is as follows:

[0082] ,

[0083] in, This indicates the number of multi-scale convolution kernels. ; This indicates that different sizes of convolution kernels are used for processing. This indicates that the kernel size is 3×3. This indicates that the kernel size is 5×5;

[0084] Spatial information Channel information Scale information The input is processed by a multi-scale attention unit to obtain multi-scale attention. The formula is expressed as follows:

[0085] ,

[0086] in, , , These represent the first trainable parameter, the second trainable parameter, and the third trainable parameter, respectively. , , It follows a standard normal distribution; This indicates the dot product operation.

[0087] 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 first convolutional layer has a 3×3 kernel size, 128 kernels, a stride of 1, and uses "same" padding. The second convolutional layer has a 3×3 kernel size, 256 kernels, a stride of 1, and uses "same" padding. The local nonlinear interaction module uses depthwise separable convolution with a 3×3×1 kernel size, a stride of 2, and "VALID" padding. The downsampling module uses the deconvolution method ConvTranspose(). The kernel size is 3×3, and the stride is set to 2, meaning that the input feature map is moved by 2 pixels during each convolution operation. This makes the size of the output feature map approximately twice that of the input, achieving an upsampling effect. The padding size is 1, meaning that zeros are added around the edges of the input feature map, so that the boundary information is preserved while the size of the output feature map increases. The dilation factor is 2, meaning that the spacing between adjacent elements in the convolution kernel is 2 pixels, which can expand the receptive field and obtain richer contextual information. A learnable bias term, bias=1, is added after the convolution operation, which helps to improve the model's expressive power and convergence speed.

[0088] The preprocessed liver data After being processed sequentially through the first convolutional layer, the second convolutional layer, the local nonlinear interaction module, and the LeakyReLU nonlinear activation function, it is then combined with spatial attention. The features are added together in the spatial dimension and then input into the downsampling module to obtain the first encoded features. ;

[0089] The second pyramid encoder consists of a third convolutional layer, a ReLU activation function, a fourth convolutional layer, a ReLU activation function, and a pooling layer. The third convolutional layer has a kernel size of 3×3, 512 kernels, a stride of 1, and uses "same" padding. The fourth convolutional layer has a kernel size of 5×5, 1024 kernels, a stride of 2, and uses "same" padding. The pooling layer uses max pooling with a pooling window size of 3×3 and a stride of 1.

[0090] The preprocessed liver data After processing by the second pyramid encoder, the second encoded feature is obtained. .

[0091] Specifically, the cross-layer connection fusion module consists 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 max pooling layer. The fifth convolutional layer has a 1×1 kernel size, 128 kernels, a stride of 1, and uses "same" padding. The average pooling layer uses average pooling with a 1×1 pooling window size and a stride of 1. The sixth convolutional layer has a 3×3 kernel size, 256 kernels, a stride of 1, and uses "same" padding. The seventh convolutional layer has a 3×3 kernel size, 512 kernels, a stride of 3, and uses "same" padding. The max pooling layer uses max pooling with a 2×2 pooling window size and a stride of 1.

[0092] For the second coding feature A pruning operation is performed, which completely removes the lower-ranked channels based on the importance index of the L1 norm, retaining only the most important feature channels, in order to simplify the model, speed up inference, and reduce storage requirements.

[0093] The L1 norm is used as a feature importance evaluation metric in the pruning operation to calculate the second encoded feature. L1 norm of the i-th feature map The formula is expressed as follows:

[0094] ,

[0095] in, and These represent the height and width of the feature map, respectively. Represents the i-th channel in the feature map The pixel value of the location; for the second encoded feature The L1 norms of the feature maps are sorted from largest to smallest, and a portion of channels are pruned from smallest to largest, with a pruning ratio of 30%. The channels corresponding to the 30% of feature maps with the smallest L1 norm after sorting are removed from the entire feature tensor, retaining only the remaining 70%. The removed channels no longer participate in subsequent forward propagation and parameter updates of the network. This yields the pruned features. ;

[0096] The first encoded feature Second coding features By adding elements one by one, the third encoded feature is obtained. The third encoding feature The input is sequentially fed into the fifth convolutional layer, the ReLU activation function, the average pooling layer, and the sixth convolutional layer for processing to obtain the fourth encoded feature. ; Pruning characteristics and fourth coding features After element-wise addition, the results are fed into the seventh convolutional layer, the ReLU activation function, and the max pooling layer to obtain the fused features. .

[0097] Specifically, the metabolic feature mapping module includes a mapping calculation layer and a high-order convolutional weighted filter;

[0098] Fusion features After standardization processing by the mapping computation layer, standardized features are obtained. The formula is expressed as follows:

[0099] ,

[0100] in, and They represent the fusion features respectively. The mean and standard deviation; then the standardized features Perform weighted adaptive normalization to obtain normalized features. The formula is expressed as follows:

[0101] ,

[0102] in, This represents the learnable weight matrix. Indicates the bias term; the normalized feature Metabolic features are obtained after processing with a high-order convolutional weighted filter. The formula is expressed as follows:

[0103] ,

[0104] in, (i,j) represents the position of the convolution kernel. The weights; Represents normalized features The pixel value at spatial location (i,j).

[0105] Specifically, the hierarchical classifier consists 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 first fully connected layer has 512 units, the second fully connected layer has 64 units, and the third fully connected layer has 4 units; the max function increments the index of the maximum value in the output of the third fully connected layer by one and then matches it with the classification dictionary for output;

[0106] The metabolic characteristics and third coding features After being added together by the flattening layers, the vector is flattened along the channel direction to obtain a one-dimensional vector, thus yielding the flattened feature. The flattening feature After passing through 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. .

[0107] Assuming the output of the third fully connected layer is [0.1, 0.3, 2.7, 0.9], then:

[0108] The maximum value is 2.7, corresponding to the index 2;

[0109] Adding 1 to the subscript 2 gives the category number 3;

[0110] If the classification dictionary is defined as {1:"Class A", 2:"Class B", 3:"Class C", 4:"Class D"}, then the output will be "Class C".

[0111] S4. Use the Adam optimizer to train and optimize the model using the loss function to obtain a well-trained model;

[0112] Specifically, the loss function adopted is the weighted cross-entropy loss function, which is expressed by the following formula:

[0113] ,

[0114] in, This represents the weighted cross-entropy loss function. Indicates the actual label, This represents the class probability predicted by the model. It is the number of categories. It is a category The weight of the class is inversely proportional to the sample frequency of that class and is the reciprocal of the sample frequency of that class, in order to balance class imbalance.

[0115] S5. After preprocessing the images to be classified in the test set, input them into the trained model to obtain the classification results.

[0116] Example 2

[0117] In this embodiment, the method of the present invention was compared with several existing advanced classification methods to verify the effectiveness and advantages of the method of the present invention.

[0118] The LiverTBT dataset was used, which contains liver pathology images of TBT-induced insulin resistance animal models. Existing methods include: CNN (Convolutional Neural Network), widely used in image classification tasks; SVM (Support Vector Machine), commonly used for few-shot learning and high-dimensional data classification; KNN (k-Nearest Neighbors), an instance-based learning method that is simple and easy to understand; and BPNN (Backpropagation Neural Network), a classic multilayer feedforward neural network. Evaluation metrics included accuracy, recall, precision, F1 score, and AUC. Each method was trained and tested on the LiverTBT dataset to evaluate its performance on the insulin resistance grading task.

[0119]

[0120] Experimental results show that our proposed method outperforms other methods in terms of accuracy, recall, precision, F1 score, and AUC (area under the ROC curve). Particularly noteworthy are our outstanding performances in AUC and F1 score, demonstrating high accuracy and robustness in insulin resistance grading tasks. Compared to CNNs, our method considers insulin resistance features more comprehensively, not just low-level image features, thus improving performance. Compared to SVMs, our method automatically learns features without requiring manual kernel function selection and performs better on complex data. Compared to KNNs, our method excels in handling high-dimensional data and is unaffected by distance metrics. Compared to BPNNs, our method, through multi-scale feature fusion and metabolic feature mapping, better captures detailed information in images. These comparative experiments demonstrate that our proposed method performs exceptionally well in insulin resistance grading tasks, outperforming existing mainstream classification methods. Its multi-scale feature fusion and metabolic feature mapping design helps to more accurately identify different levels of insulin resistance, providing strong support for clinical diagnosis.

[0121] 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 foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating insulin resistance based on multi-scale liver pathological image feature fusion, characterized in that, The method comprises the following steps: S1. Obtain the liver pathological image of a TBT-induced insulin resistance animal model, construct a LiverTBT data set, and divide it into a training set and a test set; S2. Perform data preprocessing on the liver images in the training set of the LiverTBT data set using the CLAHE algorithm to obtain preprocessed liver data; S3. Construct an insulin resistance grading evaluation model, which comprises a sliding window encoder, a multi-scale perception attention module, a first pyramid encoder, a second pyramid encoder, a cross-layer connection fusion module, a metabolic feature mapping module, and a grading classifier; and train the model using the preprocessed liver data; The sliding window encoder specifically comprises: The sliding window encoder is sequentially composed of a sliding window module, a convolution layer, and a pooling layer; the window size in the sliding window module is selected to be 32x32 pixels, the stride is selected to be 16 pixels, and the overlapping area between the windows is set to be 50%; the convolution kernel size of the convolution layer is 3x3, the number of convolution kernels is set to be 32, the ReLU activation function is used to increase the nonlinear expression capability, and the stride is set to be 1; the pooling layer is a maximum pooling layer, the pooling window size is set to be 2x2, and the stride is set to be 2; The pre-processed liver data Processed by a sliding window encoder to obtain sliding encoding features ; The multi-scale perception attention module specifically comprises: The multi-scale perception attention module comprises a spatial information processing unit, a channel information processing unit, a scale information processing unit and a multi-scale attention unit; the sliding coding features are respectively input into the spatial information processing unit, the channel information processing unit and the scale information processing unit for processing to obtain spatial information , channel information and scale information ; The formula for the processing process in the spatial information processing unit is as follows: ， wherein, represents the mean of the sliding encoding features ; represents the standard deviation of the sliding encoding features ; represents a convolution layer operation with a kernel size of 3x3 and a number of kernels set to 64; represents a RELU activation function; represents a Sigmoid activation function; The formula for the processing process in the channel information processing unit is as follows: ， wherein, represents a constant, ; represents an element-wise multiplication, represents an averaging operation over spatial dimensions, represents an operation along the 1st dimension and the 2nd dimension; The formula for the processing process in the scale information processing unit is as follows: ， wherein, represents the number of multi-scale convolution kernels, ; represents processing with different sizes of convolution kernels, represents a convolution kernel size of 3x3, represents a convolution kernel size of 5x5; The spatial information , channel information , scale information is input to the multi-scale attention unit for processing to obtain multi-scale attention , which is expressed by the following formula: ， wherein, , , respectively represent a first trainable parameter, a second trainable parameter, and a third trainable parameter; , , obeys a standard normal distribution; denotes a dot product operation; The first pyramid encoder specifically comprises: The first pyramid encoder comprises a first convolution layer, a second convolution layer, a local nonlinear interaction module, a LeakyReLU nonlinear activation function, and a down-sampling module; the convolution kernel size of the first convolution layer is 3x3, the number of convolution kernels is 128, the stride is 1, and the "same" padding is used; the convolution kernel size of the second convolution layer is 3x3, the number of convolution kernels is 256, the stride is 1, and the "same" padding is used; the local nonlinear interaction module uses a depth separable convolution, the convolution kernel size is 3x3x1, the stride is 2, and the "VALID" padding mode is used; the down-sampling module uses a deconvolution method, the convolution kernel size is 3x3, the stride is set to be 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 sequentially processed by a first convolutional layer, a second convolutional layer, a local nonlinear interaction module, and a LeakyReLU nonlinear activation function, the first encoding feature is added in the spatial dimension to an input into a down-sampling module After being sequentially processed by a first convolutional layer, a second convolutional layer, a local nonlinear interaction module, and a LeakyReLU nonlinear activation function, the first encoding feature is added in the spatial dimension to an input into a down-sampling module ; The second pyramid encoder specifically comprises: The second pyramid encoder is sequentially composed of a third convolution layer, a ReLU activation function, a fourth convolution layer, a ReLU activation function, and a pooling layer; the convolution kernel size of the third convolution layer is 3x3, the number of convolution kernels is 512, the stride is 1, and the "same" padding is used; the convolution kernel size of the fourth convolution layer is 5x5, the number of convolution kernels is 1024, the stride is 2, and the "same" padding is used; the pooling layer adopts maximum pooling, the pooling window size is 3x3, and the stride is 1; The pre-processed liver data processed by a second pyramid encoder to obtain second encoded features ; The cross-layer connection fusion module specifically comprises: The cross-layer connection fusion module is sequentially 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; the fifth convolutional layer has a kernel size of 1*1, a kernel number of 128, a stride of 1, and uses "same" padding; the average pooling layer adopts average pooling with a pooling window size of 1*1 and a stride of 1; the sixth convolutional layer has a kernel size of 3*3, a kernel number of 256, a stride of 1, and uses "same" padding; the seventh convolutional layer has a kernel size of 3*3, a kernel number of 512, a stride of 3, and uses "same" padding; the maximum pooling layer adopts maximum pooling with a pooling window size of 2*2 and a stride of 1; to the second encoding feature pruning operation, wherein the L1 norm is used as an importance evaluation index of the feature, to calculate the second encoding feature the L1 norm of the i-th feature map in the second encoding feature The formula is as follows: ， wherein, and respectively represent the height and width of the feature map, represents the pixel value of the i-th channel in the feature map at the position the second encoded feature The L1 norm of the feature map in the second encoded feature is sorted from large to small, and a part of the channels is pruned from small to large, and the pruning ratio is set to 30%, to obtain the pruned feature ; The first encoding feature and the second encoding feature are added element by element to obtain a third encoding feature ; the third encoding feature is sequentially input to a fifth convolutional layer, a ReLU activation function, an average pooling layer and a sixth convolutional layer for processing to obtain a fourth encoding feature ; the pruning feature and the fourth encoding feature are added element by element and then input to a seventh convolutional layer, a ReLU activation function and a maximum pooling layer to obtain a fusion feature ; The metabolic feature mapping module specifically includes: The metabolic feature mapping module includes a mapping calculation layer and a high-order convolution weighted filter; Fusion features The standardized features are obtained by performing standardization processing through the mapping calculation layer The formula is as follows: ， wherein, and respectively represent the mean and standard deviation of the fused features ; then a weighted adaptive normalization operation is performed on the normalized features to obtain normalized features , which is expressed by the following formula: ， wherein, denotes a learnable weight matrix, denotes a bias term; the normalized features are processed through a high order convolutional weighting filter to obtain metabolic features , which is expressed as follows: ， wherein, (i,j) is the weight of the convolution kernel at position (i,j) is the weight of the convolution kernel at position represents the normalized feature pixel value at spatial position (i,j); The hierarchical classifier specifically includes: The hierarchical classifier is sequentially composed of a flattening layer, a first full connection layer, a second full connection layer, a third full connection layer, and a max function; the first full connection layer has a unit number of 512, the second full connection layer has a unit number of 64, and the third full connection layer has a unit number of 4; the max function matches the output of the third full connection layer with a classification table dictionary after adding one to the subscript corresponding to the maximum value. The metabolic features And the third encoding features After element-by-element addition through the flattening layer, the feature is flattened into a one-dimensional vector in the channel direction to obtain a flattened feature ; The flattened feature After the first fully connected layer, the second fully connected layer, the third fully connected layer and the max function, the insulin resistance classification is obtained ; S4. The model is trained and optimized by using an Adam optimizer through a loss function, to obtain a trained model. S5. The hierarchical result is obtained by inputting the preprocessed image to be classified in the test set into the trained model.

2. The method for evaluating insulin resistance based on multi-scale liver pathological image feature fusion according to claim 1, characterized in that, In step S1, the insulin resistance level of the liver pathological image is labeled, and is 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% is used as a test set.

3. The method for insulin resistance evaluation based on multi-scale liver pathological 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 block size Tile Size is set to 8x8, indicating that the original liver pathological image is divided into 8 rows and 8 columns, a total of 64 small regions; the clipping threshold or the limited contrast parameter Clip Limit is set to 2.0; the interpolation processing process adopts a bilinear interpolation method; all images in the training set are processed by the CLAHE algorithm to obtain the preprocessed liver data .

4. The method for insulin resistance evaluation based on multi-scale liver pathological image feature fusion according to claim 3, characterized in that, In step S4, the loss function adopts a weighted cross-entropy loss function, and the formula is as follows: ， wherein, represents a weighted cross-entropy loss function, represents an actual label, represents a class probability predicted by a model; represents a number of classes, ; represents a weight of a class .

Images