Kidney lesion image classification system based on deep learning
By designing an equalization loss function and constructing a domain weight matrix, the model parameters were optimized, which solved the problems of noise robustness and lesion edge segmentation accuracy in the kidney image classification system, thus improving the reliability and accuracy of kidney image classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE FIRST MEDICAL CENT CHINESE PLA GENERAL HOSPITAL
- Filing Date
- 2026-04-17
- Publication Date
- 2026-05-15
AI Technical Summary
Existing kidney image classification systems are not robust to noise and artifacts, have low accuracy in lesion edge segmentation, and blurry mask boundaries, resulting in low reliability of the final classification. They also ignore the domain-specific dependencies of lesion pixels, overfit to image noise, mask the features of small lesions, and lack characterization of the continuity of lesion direction, leading to poor classification results.
We design an equalization loss function based on modulation factor and gradient penalty loss, construct a direction-aware domain weight matrix, optimize the model through gradient penalty loss and parameter adjustment, introduce a domain delay feature term, strengthen the feature weight of small lesions, reduce noise sensitivity, and characterize the regional features of lesions.
It improves the reliability and accuracy of kidney image classification, reduces sensitivity to noise and artifacts, enhances the segmentation accuracy and classification effect of small lesions, and reduces misjudgments.
Smart Images
Figure CN122049546A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing, specifically to a deep learning-based image classification system for kidney lesions. Background Technology
[0002] Kidney image classification systems are auxiliary systems based on medical image processing and artificial intelligence technologies that automatically extract lesion-related features from kidney images to accurately identify and classify kidney image types. However, general kidney image classification systems suffer from poor robustness to noise and artifacts in kidney images, low accuracy in lesion edge segmentation, and blurred mask boundaries, leading to low reliability of the final classification. Furthermore, they often ignore the domain-specific dependencies of lesion pixels, overfit to image noise, mask subtle lesion features, and lack characterization of the continuity of lesion direction, resulting in poor classification performance. Summary of the Invention
[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a deep learning-based kidney lesion image classification system. Addressing the problems of poor robustness to noise and artifacts, low lesion edge segmentation accuracy, and blurred mask boundaries in general kidney image classification systems, leading to low final classification reliability, this solution employs a modulation factor-based equalization loss design to force the model to focus on small lesions and lesion regions with blurred edges, thereby improving the segmentation accuracy of edge regions. A gradient-based penalty loss design is also implemented, introducing two parameters to separately control the penalty for false positives and false negatives. Parameter updates are performed based on gradients under extreme offset conditions, reducing sensitivity to noise and artifacts, thus improving the final reliability of kidney image classification. This system is designed to address the shortcomings of general kidney lesion image classification systems. Traditional kidney image classification systems suffer from problems such as ignoring the domain-specific dependencies of lesion pixels, overfitting image noise, masking the features of minute lesions, and lacking characterization of the directional continuity of lesions, leading to poor classification results. This solution constructs a direction-aware domain-specific weight matrix to characterize the domain-specific adjacency relationships between superpixel blocks, reducing misclassifications caused by ignoring directional features. It introduces a surrounding domain-specific delay feature term, incorporating the features of all surrounding superpixel blocks into the classification calculation, reflecting the domain-specific continuity of lesions. By using a uniform penalty term, it weakens the weight of noise features and strengthens the feature weight of minute lesions, aligning with the regional characteristics of kidney lesions and avoiding noise dominating the classification results, thereby improving the classification performance of kidney images.
[0004] The technical solution adopted by the present invention is as follows: The deep learning-based kidney lesion image classification system provided by the present invention includes a kidney image acquisition module, a kidney lesion segmentation model construction module, a model parameter adjustment module, a kidney image classification model design module, and a kidney image classification module;
[0005] The kidney image acquisition module acquires historical kidney image data and performs preprocessing and annotation to construct a kidney image segmentation set.
[0006] The kidney lesion segmentation model construction module is based on a kidney image segmentation set. It constructs a kidney lesion segmentation model by building an encoder-decoder end-to-end network and a loss function based on gradient penalty.
[0007] The model parameter control module optimizes and controls the parameters of the kidney lesion segmentation model based on parameter offset;
[0008] The kidney image classification model design module integrates the lesion mask features output by the kidney lesion segmentation model, constructs a domain weight matrix, and designs a kidney image classification model.
[0009] The kidney image classification module classifies real-time kidney image data based on a kidney lesion segmentation model and a kidney image classification model.
[0010] Furthermore, the kidney image acquisition module acquires historical kidney image data, performs preprocessing including ROI cropping, grayscale normalization, image resampling, and Gaussian denoising; and performs image annotation, completing segmentation annotation and classification label value assignment; and constructs a kidney image segmentation set.
[0011] Furthermore, the kidney lesion segmentation model construction module specifically includes:
[0012] Model network architecture construction: Construct an end-to-end segmentation network of encoder-decoder. The encoder adopts a deep convolutional neural network to extract multi-scale features of kidney images. The decoder adopts an upsampling + convolution structure to gradually restore image resolution. At the same time, domain-specific batch normalization is added to each decoding layer. The output layer adopts the softmax activation function to output a pixel-level predicted mask of kidney lesions.
[0013] Loss function design: A balanced loss is designed to reduce the loss weight of classified samples through a dual mechanism of class balancing and modulation factor; a gradient penalty loss is designed, which introduces two parameters to control the penalty intensity for false positives and false negatives respectively; finally, the total loss function of the kidney lesion segmentation model is obtained.
[0014] Furthermore, the model parameter adjustment module specifically includes:
[0015] Define the core optimization objective of parameter tuning, find extreme cost cases around the model parameters and achieve global minimization to guide model convergence;
[0016] Calculate the optimal parameter offset vector to ensure training efficiency;
[0017] Define the parameter update rules for the kidney lesion segmentation model, and complete the parameter update based on the gradient of extreme offset cases.
[0018] Furthermore, the kidney image classification model design module specifically includes:
[0019] The renal lesion domain weight matrix is constructed as follows: The pixel-level predicted mask output by the segmentation model is multiplied pixel-by-pixel with the renal image to obtain a masked renal image; Using the masked renal image as input, the encoder of the segmentation model is called as the feature extraction backbone network to extract multi-scale lesion depth features. Channel fusion and dimensionality reduction of multi-scale features are completed through 1×1 convolution to obtain a high-dimensional feature matrix of a single image; The masked renal image is divided into non-overlapping superpixel blocks, and each superpixel block corresponds to a feature subset in the feature matrix; Finally, the renal lesion domain weight matrix is constructed, and orientation-aware domain weights are introduced.
[0020] The classification output design incorporates lesion domain delay feature terms, integrates the surrounding lesion features of superpixel blocks into the classification model, quantifies the lesion feature dependencies of superpixels at different domain locations, and constructs a classification model adapted to the distribution characteristics of renal lesion domains; thus obtaining the classification output of the entire image.
[0021] Classification loss estimation: Based on multi-class cross-entropy loss, a lesion feature smoothing penalty term based on the second derivative is added to control the roughness of the domain interdependence function and feature weight function, thereby minimizing the classification loss;
[0022] Model training: When training the classification model, the parameter control logic of the previous kidney lesion segmentation model is completely followed, and the model parameter peripheral offset optimization is introduced in the parameter update step.
[0023] Furthermore, the kidney image classification module classifies the real-time acquired kidney image data based on the kidney lesion segmentation model and the kidney image classification model; it performs preprocessing on the real-time acquired kidney images, inputs the trained kidney lesion segmentation model, and outputs pixel-level lesion masks; then it fuses the original image with the lesion mask to generate a masked kidney image, inputs it into the kidney image classification model, and obtains the final classification result of the kidney image.
[0024] The beneficial effects achieved by the present invention using the above solution are as follows:
[0025] (1) To address the problems of poor robustness to noise and artifacts in general kidney image classification systems, low lesion edge segmentation accuracy, and blurred mask boundaries, which lead to low final classification reliability, this scheme designs an equalization loss based on modulation factor, forces the model to focus on small lesions and lesion regions with blurred edges, and improves the segmentation accuracy of edge regions; designs a gradient penalty loss, and introduces two parameters to adjust the penalty intensity for false detection and false negative detection respectively; completes parameter updates based on gradients in extreme offset cases, reduces sensitivity to noise and artifacts, and thus improves the final kidney image classification reliability.
[0026] (2) To address the problems of general kidney image classification systems, such as ignoring the domain-location dependency of lesion pixels, overfitting image noise, masking the features of small lesions, and lacking the characterization of the directional continuity of lesions, which leads to poor classification results, this scheme constructs a direction-aware domain-location weight matrix to characterize the domain-location adjacency relationship between superpixel blocks, thereby reducing misjudgments caused by ignoring directional features; it introduces a surrounding domain-location delay feature term to incorporate the features of all surrounding superpixel blocks of the current superpixel block into the classification calculation, reflecting the domain-location continuity of the lesion; and it weakens the weight of noise features and strengthens the feature weight of small lesions through a uniform penalty term, which fits the regional features of kidney lesions and avoids noise dominating the classification results; thereby improving the kidney image classification effect. Attached Figure Description
[0027] Figure 1 This is a flowchart illustrating the deep learning-based kidney lesion image classification system provided by the present invention.
[0028] 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. Detailed Implementation
[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0030] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0031] Example 1, see Figure 1 The present invention provides a deep learning-based kidney lesion image classification system, including a kidney image acquisition module, a kidney lesion segmentation model construction module, a model parameter adjustment module, a kidney image classification model design module, and a kidney image classification module;
[0032] The kidney image acquisition module acquires historical kidney image data, performs preprocessing and annotation, constructs a kidney image segmentation set, and sends the data to the kidney lesion segmentation model construction module.
[0033] The kidney lesion segmentation model construction module is based on a kidney image segmentation set. It constructs a kidney lesion segmentation model by building an end-to-end encoder-decoder network and a loss function based on gradient penalty; and sends the data to the model parameter adjustment module.
[0034] The model parameter adjustment module optimizes and adjusts the parameters of the kidney lesion segmentation model based on parameter offset; and sends the data to the kidney image classification model design module.
[0035] The kidney image classification model design module integrates the lesion mask features output by the kidney lesion segmentation model, constructs a domain weight matrix, designs a kidney image classification model, and sends the data to the kidney image classification module.
[0036] The kidney image classification module classifies real-time kidney image data based on a kidney lesion segmentation model and a kidney image classification model.
[0037] Example 2, see Figure 1 This embodiment is based on the above embodiment. The kidney image acquisition module acquires historical kidney image data (CT images and MRI images), performs preprocessing, including ROI cropping, grayscale normalization, image resampling, and Gaussian denoising; and performs image annotation, completing segmentation annotation and classification label value assignment; the segmentation annotation includes background / normal renal parenchyma / lesion area; the classification label includes stones, cysts, tumors, and normal; and constructs a kidney image segmentation set.
[0038] Example 3, see Figure 1 This embodiment is based on the above embodiment. The kidney lesion segmentation model construction module is based on a kidney image segmentation set. Through model network architecture construction, multi-loss function collaborative design, and model parameter adjustment and optimization, a kidney lesion segmentation model is designed; specifically, it includes:
[0039] Model network architecture construction: An end-to-end segmentation network consisting of an encoder and decoder is constructed. The encoder uses a deep convolutional neural network (ResNet50) to extract multi-scale features from kidney images, from low-level edge and texture features to high-level kidney anatomical structures and lesion semantic features, with a focus on enhancing the extraction of lesion edge features. The feature fusion layer fuses the shallow, mid-level, and high-level features extracted by the encoder at multiple scales, amplifying the feature signals of small lesions and solving the problem that features of small kidney lesions are easily obscured. The decoder adopts an upsampling + convolution structure to gradually restore image resolution, while adding domain-specific batch normalization to each decoding layer. The output layer uses a softmax activation function to output a pixel-level predicted mask of kidney lesions (background / normal renal parenchyma / lesion area).
[0040] Loss function design; defining the total loss function for the kidney lesion segmentation model. , represented as: ;in, , , and These are the cross-entropy loss, Dice loss, equilibrium loss, and gradient penalty loss for the original samples, respectively. An equilibrium loss is designed by using a dual mechanism of class balancing and modulation factors to reduce the loss weight of classified samples, forcing the model to focus on the blurred edges of kidney lesions and the difficult-to-segment samples of small lesions, expressed as: A gradient penalty loss is designed, introducing two parameters to control the penalty intensity for false positives (normal renal parenchyma misdiagnosed as lesions) and false negatives (lesions misdiagnosed as normal renal parenchyma), expressed as: ;in, This is the category balancing weight, with a value ranging from 0.2 to 0.8; and , respectively, are the true label and lesion prediction mask for the i-th pixel belonging to class c. The true label has values of 0 (background), 1 (normal renal parenchyma), and 2 (lesion region). The lesion prediction mask has values of 0 to 1, representing the probability that the pixel belongs to a class. γ is the attention intensity parameter. The larger the value, the stronger the suppression of loss of easily classified samples such as normal renal parenchyma and background. The more the model focuses on difficult-to-segment lesion regions, and the value ranges from 1 to 5. N is the total number of pixels in the image. It is a false positive penalty factor, with a value of 0.1~0.5, which inhibits the misdiagnosis of normal renal parenchyma / background as renal lesions and reduces the risk of clinical overdiagnosis; This is the total number of falsely detected pixels; It is a penalty factor for missed detection, with a value of 0.5~1.0, which inhibits the situation where kidney lesions are misdiagnosed as normal renal parenchyma / background; This represents the total number of missed pixels. Prioritize reducing the rate of missed lesion detection;
[0041] The equalization loss allows the model to actively focus on the blurred edges of kidney lesions and difficult-to-segment samples such as small cysts / stones, improving the segmentation accuracy of lesion edges; the gradient penalty loss aligns with the clinical diagnostic needs of kidney lesions, prioritizing the reduction of lesion false negatives, which is in line with the clinical value orientation of medical image segmentation.
[0042] Example 4, see Figure 1 This embodiment is based on the above embodiment. The model parameter adjustment module is designed to avoid the segmentation model's parameters being sensitive to noise in kidney images. It optimizes the parameters by considering extreme cases around the parameters, reducing parameter sensitivity and improving the segmentation model's generalization ability on kidney images from clinically unknown devices. Specifically, it includes:
[0043] Define the core optimization objective of parameter tuning: find extreme cost cases around the model parameters and achieve global minimization to guide model convergence, expressed as: ;in, These are all the trainable parameters of the kidney lesion segmentation model, including the encoder, decoder, domain-specific batch normalized weights and biases; It is the model parameter offset vector, which represents the small adjustment values of the parameters; It is the L2 norm; It is the radius around the parameter, with a value of 1e. -6 ~1e -4 The maximum magnitude of the control parameter offset; It is the total loss function of the kidney lesion segmentation model;
[0044] To ensure training efficiency, the optimal parameter offset vector is calculated, as follows: ;in, It is the optimal parameter offset vector, which is the standardized offset in the direction of the fastest loss growth, realizing the extreme cost situation around the parameters; The total loss of the kidney lesion segmentation model is relative to the parameters. The gradient points in the direction of the fastest growth of the total loss;
[0045] The parameter update rule for the kidney lesion segmentation model is defined, and the parameter update is completed based on the gradient of extreme offset cases, ensuring that the segmentation loss of all parameters in the surrounding area remains at a low level after the parameter update, expressed as: ; These are all trainable parameters of the updated kidney lesion segmentation model; This is the model learning rate, with a value of 1e. -5 ~1e -3 Use the Adam optimizer;
[0046] Minimizing the loss value and loss steepness allows the model to converge, preventing small changes in parameters from causing large fluctuations in the segmentation loss and improving the model's robustness to noise and artifacts in kidney images.
[0047] By performing the above operations, this scheme addresses the problems of poor robustness to noise and artifacts, low lesion edge segmentation accuracy, and blurred mask boundaries in general kidney image classification systems, which lead to low final classification reliability. It designs an equalization loss based on a modulation factor to force the model to focus on small lesions and lesion regions with blurred edges, thus improving the segmentation accuracy of edge regions. It also designs a gradient-based penalty loss, introducing two parameters to separately control the penalty for false positives and false negatives. Parameter updates are completed based on gradients under extreme offset conditions, reducing sensitivity to noise and artifacts, thereby improving the final reliability of kidney image classification.
[0048] Example 5, see Figure 1 This embodiment is based on the above embodiment. The kidney image classification model design module is designed to avoid ignoring the domain-specific dependencies between pixels (the surrounding associations of pixels in the lesion area and the domain-specific continuity of normal kidney parenchyma). It integrates the lesion mask features output by the previous segmentation model, introduces pixel domain-specific delay feature terms into the classification model, and fuses grayscale, texture, and lesion mask features to construct a kidney image classification model adapted to the domain-specific structure of kidney images; specifically including:
[0049] The renal lesion region weight matrix is constructed. The pixel-level predicted mask output by the segmentation model is multiplied pixel-by-pixel with the renal image to obtain a masked renal image, completely removing the background region and retaining only the effective pixels of the normal renal parenchyma and lesion areas. Using the masked renal image as input, the encoder of the segmentation model is called as the feature extraction backbone network to extract multi-scale lesion depth features: the shallow layer extracts lesion edges and texture features (high grayscale edges of stones, smooth boundaries of cysts); the middle layer extracts lesion morphology and contour features (irregular shapes of tumors, round features of stones); and the high layer extracts lesion semantic and anatomical structural features (cysts in the renal cortex). The distribution of tumors and their infiltration characteristics in the renal medulla are analyzed. Multi-scale feature fusion and dimensionality reduction are achieved through 1×1 convolution to obtain a high-dimensional feature matrix of a single image, denoted as F. The masked kidney image is divided into S (values 16~64) non-overlapping superpixel blocks, each corresponding to a feature subset in the feature matrix F, achieving feature localization and allocation. Finally, a kidney lesion local weight matrix is constructed to characterize the local adjacency relationship between superpixel blocks. Introducing direction-aware local weights improves classification accuracy for irregular tumors extending along a certain direction, reducing misclassifications caused by ignoring directional continuity. Matrix elements... , It is the Euclidean distance between superpixel blocks m and n; It is the direction sensitivity coefficient, with a value of 1 to 5; It is the geometric orientation angle of the superpixel block m→n; is the major axis direction angle of the kidney as a whole in the i-th image, and is the direction of the first principal component obtained by principal component analysis of the kidney contour;
[0050] The classification output design incorporates lesion location delay feature terms, integrating the surrounding lesion features of superpixel blocks into the classification model. This quantifies the lesion feature dependencies of superpixels at different location positions, constructing a classification model adapted to the location distribution characteristics of kidney lesions. The core principle is to quantify the location dependency strength of lesion features through location interdependence functions and to characterize the contribution of different lesion features to the classification results through feature weight functions, expressed as: ;in, is the classification prediction scalar value of the m-th superpixel block corresponding to the t-th class label in the i-th kidney image; S is the total number of superpixel blocks, and p is the superpixel block index; is the element of the domain weight matrix, which describes the domain dependence of kidney lesions between superpixel blocks m and p; D is the total dimension of discrete high-dimensional lesion features in the kidney image, d is the lesion feature dimension index; t is the classification label dimension index. It is a bivariate lesion domain location interdependence function that quantifies the domain location dependence strength of the t-th dimension classification label and the d-th dimension feature of kidney lesions; It is a lesion feature weight function, which characterizes the degree of contribution of the d-th lesion feature to the t-th classification label; It is the bit field of the image pixels; It is a random error term that follows a normal distribution with a mean of 0 and a constant variance; and These are the d-th dimension lesion features of superpixel block p and the d-th dimension lesion features of m in the i-th kidney image, respectively. and As a hyperparameter, it participates in model parameter tuning and is initialized using Xavier normality;
[0051] It is the surrounding domain location delay feature term, which integrates the features of all surrounding superpixel blocks of the current superpixel block to reflect the domain location continuity of the lesion; It is the current superpixel's own feature term, which describes the contribution of its own lesion features to the classification label; the two terms are superimposed to obtain the classification label prediction value of superpixel block m;
[0052] The classification output of the entire image is represented as follows: ;in, It is the classification prediction result of the entire image of the i-th kidney image; It is the softmax activation function; It is a classification prediction vector, consisting of 4 The results are obtained by integrating the data; the label corresponding to the highest probability is used as the classification result.
[0053] Classification loss estimation: Kidney images suffer from imaging noise and the tendency for small lesion features to be submerged. To avoid overfitting to local noise, a multi-class cross-entropy loss is used as a basis, with an added lesion feature smoothing penalty term based on the second derivative. This term controls the roughness of the domain interdependence function and the feature weight function, minimizing the classification loss. Specifically, the smoothing penalty term of the domain interdependence function... Represented as: ;in, It is the pixel domain and bit domain of the kidney image; It is the current pixel coordinate; yes The second-order partial derivatives at the corresponding pixel coordinates; to avoid the model over-capturing spurious domain correlations caused by local imaging noise, ensuring the continuity of the domain dependence of kidney lesions; the uniformity penalty term of the feature weight function. Represented as: ; yes The second-order partial derivative at the corresponding pixel coordinates; to avoid noise features dominating the classification results, to fit the regional characteristics of kidney lesions, and to strengthen the weight of small lesion features; the final loss function L is expressed as: ; It is cross-entropy loss; and This is the penalty weighting coefficient, with a value ranging from 0.001 to 0.1;
[0054] Model training: During the training of the classification model, the parameter control logic of the previous kidney lesion segmentation model is completely followed. In the parameter update step, the model parameter periphery offset optimization is introduced to ensure the generalization of the classification model to kidney images, while improving the model's robustness to imaging noise and artifacts.
[0055] By performing the above operations, this scheme addresses the problems of general kidney image classification systems, such as ignoring the domain-specific dependencies of lesion pixels, overfitting image noise, masking the features of minor lesions, and lacking characterization of the directional continuity of lesions, which leads to poor classification results. It constructs a direction-aware domain-specific weight matrix to characterize the domain-specific adjacency relationships between superpixel blocks, reducing misclassifications caused by ignoring directional features. It introduces a surrounding domain-specific delay feature term, incorporating the features of all surrounding superpixel blocks into the classification calculation, reflecting the domain-specific continuity of the lesion. Furthermore, it weakens the weight of noise features and strengthens the feature weight of minor lesions through a uniform penalty term, aligning with the regional characteristics of kidney lesions and preventing noise from dominating the classification results. This ultimately improves the kidney image classification performance.
[0056] Example 6, see Figure 1This embodiment is based on the above embodiment. The kidney image classification module classifies the real-time acquired kidney image data based on the kidney lesion segmentation model and the kidney image classification model. It performs preprocessing on the real-time acquired kidney images, inputs the trained kidney lesion segmentation model, and outputs pixel-level lesion masks. Then, it fuses the original image with the lesion mask to generate a masked kidney image, inputs it into the kidney image classification model, and obtains the final classification result of the kidney image.
[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0058] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A deep learning-based image classification system for kidney lesions, characterized in that: The system includes a kidney image acquisition module, a kidney lesion segmentation model construction module, a model parameter adjustment module, a kidney image classification model design module, and a kidney image classification module; The kidney image acquisition module acquires historical kidney image data and performs preprocessing and annotation to construct a kidney image segmentation set. The kidney lesion segmentation model construction module is based on a kidney image segmentation set. It constructs a kidney lesion segmentation model by building an encoder-decoder end-to-end network and a loss function based on gradient penalty. The model parameter control module optimizes and controls the parameters of the kidney lesion segmentation model based on parameter offset; The kidney image classification model design module integrates the lesion mask features output by the kidney lesion segmentation model, constructs a domain weight matrix, and designs a kidney image classification model. The kidney image classification module classifies real-time kidney image data based on a kidney lesion segmentation model and a kidney image classification model.
2. The deep learning-based kidney lesion image classification system according to claim 1, characterized in that: The kidney lesion segmentation model construction module specifically includes: Model network architecture construction: Construct an end-to-end segmentation network of encoder-decoder. The encoder adopts a deep convolutional neural network to extract multi-scale features of kidney images. The decoder adopts an upsampling + convolution structure to gradually restore image resolution. At the same time, domain-specific batch normalization is added to each decoding layer. The output layer adopts the softmax activation function to output a pixel-level predicted mask of kidney lesions. Loss function design: A balanced loss is designed to reduce the loss weight of easily classified samples through a dual mechanism of class balancing and modulation factor; a gradient penalty loss is designed, which introduces two parameters to control the penalty intensity for false positives and false negatives respectively; finally, the total loss function of the kidney lesion segmentation model is obtained.
3. The deep learning-based kidney lesion image classification system according to claim 2, characterized in that: The model parameter adjustment module specifically includes: Define the core optimization objective of parameter tuning, find extreme cost cases around the model parameters and achieve global minimization to guide model convergence; Calculate the optimal parameter offset vector to ensure training efficiency; Define the parameter update rules for the kidney lesion segmentation model, and complete the parameter update based on the gradient of extreme offset cases.
4. The deep learning-based kidney lesion image classification system according to claim 3, characterized in that: The kidney image classification model design module specifically includes: The renal lesion domain weight matrix is constructed as follows: The pixel-level predicted mask output by the segmentation model is multiplied pixel-by-pixel with the renal image to obtain a masked renal image; Using the masked renal image as input, the encoder of the segmentation model is called as the feature extraction backbone network to extract multi-scale lesion depth features. Channel fusion and dimensionality reduction of multi-scale features are completed through 1×1 convolution to obtain a high-dimensional feature matrix of a single image; The masked renal image is divided into non-overlapping superpixel blocks, and each superpixel block corresponds to a feature subset in the feature matrix; Finally, the renal lesion domain weight matrix is constructed, and orientation-aware domain weights are introduced. The classification output design incorporates lesion domain delay feature terms, integrates the surrounding lesion features of superpixel blocks into the classification model, quantifies the lesion feature dependencies of superpixels at different domain locations, and constructs a classification model adapted to the distribution characteristics of kidney lesion domains; thus obtaining the classification output of the entire image. Classification loss estimation; Model training: When training the classification model, the parameter control logic of the previous kidney lesion segmentation model is completely followed, and the model parameter peripheral offset optimization is introduced in the parameter update step.
5. The deep learning-based kidney lesion image classification system according to claim 4, characterized in that: The classification loss estimation is based on multi-class cross-entropy loss, with the addition of a lesion feature uniformity penalty term based on the second derivative to control the roughness of the domain interdependence function and feature weight function, thereby minimizing the classification loss.
6. The deep learning-based kidney lesion image classification system according to claim 5, characterized in that: The kidney image acquisition module acquires historical kidney image data, performs preprocessing including ROI cropping, grayscale normalization, image resampling, and Gaussian denoising; and performs image annotation, completing segmentation annotation and classification label value assignment; and constructs a kidney image segmentation set.
7. The deep learning-based kidney lesion image classification system according to claim 6, characterized in that: The kidney image classification module classifies real-time acquired kidney image data based on a kidney lesion segmentation model and a kidney image classification model; it performs preprocessing on real-time acquired kidney images, inputs the trained kidney lesion segmentation model, and outputs pixel-level lesion masks. The original image is then fused with a lesion mask to generate a masked kidney image, which is then input into a kidney image classification model to obtain the final classification result of the kidney image.