Medical image classification method based on curriculum-type enhancement scheduling and hybrid enhancement

Abstract

The invention discloses a medical image classification method based on curriculum-type enhancement scheduling and hybrid enhancement. The method comprises the following steps: S100, constructing a curriculum-type data enhancement mechanism; s200, performing a staged hybrid enhancement strategy; s300, carrying out multi-stage model training and optimization; and S400, carrying out model interpretability output generation. According to the method, the stability and generalization performance of the model in the multi-stage training process are improved, the interpretability of a model output result is enhanced, and reliable technical support is provided for a clinical auxiliary diagnosis system.

Description

A Medical Image Classification Method Based on Curriculum-Based Augmentation Scheduling and Hybrid Augmentation Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical image analysis technology, specifically to a medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence technology, deep learning models, especially convolutional neural networks and visual Transformers, have achieved remarkable results in the field of medical image analysis. These technologies have shown great potential in image recognition tasks for various diseases such as pneumonia, tuberculosis, and tumors. However, in actual medical image classification applications, existing technologies still face many challenges and limitations: (1) The singularity and static nature of data augmentation strategies. Most current medical image classification methods adopt fixed data augmentation methods, such as random rotation, flipping, brightness adjustment, etc., and remain unchanged throughout the training process. This static augmentation strategy has obvious defects: in the early stage of model training, overly complex or strong augmentation may lead to feature learning chaos and affect the stable convergence of the model; while in the later stage of training, fixed augmentation strategies may not be able to provide sufficient data diversity, resulting in insufficient model generalization ability and difficulty in adapting to the changing imaging conditions in actual applications.

[0003] (2) Regularization techniques such as non-adaptive mixing enhancement (e.g., Mixup) and label smoothing typically use fixed parameter settings throughout the training phase. However, the learning capacity and needs of the model differ at different training phases. Mixup with fixed parameters may lead to feature confusion due to excessive mixing in the later stages of training, while label smoothing with a fixed degree may not effectively prevent overfitting in the later stages of training. Existing techniques lack the ability to dynamically adjust these regularization parameters according to the training progress.

[0004] (3) Special Issues of Medical Imaging Datasets Medical imaging datasets often contain multiple examinations or images from different angles of the same patient. Randomly assigning these samples from the same patient to the training and validation sets can lead to serious information leakage, resulting in an inflated performance evaluation of the model on the validation set, which fails to truly reflect its generalization ability in practical applications. This problem is particularly prominent in medical image analysis, but existing methods often overlook or fail to address it properly.

[0005] (4) Lack of interpretability of model decisions In clinical practice, doctors not only need the model to provide accurate classification results, but also need to understand the basis on which the model makes the judgment. Traditional deep learning classification models usually only output class probabilities and lack an intuitive and visual explanation of the decision-making process. This makes it difficult for clinicians to understand and trust the model's judgment results, which limits the practical application of intelligent diagnostic systems in the clinical environment. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation, which not only improves the stability and generalization performance of the model in the multi-stage training process, but also enhances the interpretability of the model output results, providing reliable technical support for clinical auxiliary diagnostic systems.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows.

[0008] A medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation includes the following steps: S100. Construction of curriculum-based data augmentation mechanism: Load the medical image dataset, and use group-aware hierarchical cross-validation to divide the dataset into training and validation sets, ensuring that images of the same patient do not appear in the training and validation sets simultaneously; divide the model training process into three curriculum stages executed sequentially with increasing difficulty according to a predetermined number of rounds; S200. Implementation of phased hybrid augmentation strategy: Apply data augmentation strategies with different intensities and data distributions in different curriculum stages, and the complexity and regularization strength of the augmentation strategy change with the progression of the stages; S300. Multi-stage model training and optimization: Dynamically adjust the augmentation strategy parameters in each of the divided curriculum stages, and use a multimodal regularization strategy to train and optimize the backbone network of the model; S400. Generation of interpretable model output: Use the trained model to classify medical images, and generate heatmaps based on the gradient information output by the model, which are overlaid on the original images to visualize the lesion regions on which the model's decisions are based.

[0009] Preferably, in step S100, group-aware hierarchical cross-validation is implemented in the following way: extracting the stem value of the medical image sample file name as the patient identifier; grouping the samples based on the stem value so that the images of the same patient belong to the same group; and maintaining the integrity of the samples within the group when dividing the training set and the validation set, ensuring that the samples of the same group appear only in the training set or the validation set.

[0010] Preferably, in step S100, the model training process is divided into three course stages executed sequentially and with increasing difficulty according to a predetermined number of rounds. Specifically, this includes: S110. Sample complexity assessment: Based on the image feature entropy, edge clarity, and lesion salience indicators, the training samples are assessed for difficulty and stratified; S120. Stage division and strategy configuration: Based on the difficulty assessment results, the model training process is divided into a basic stage, a medium stage, and a high-difficulty stage, and enhancement types and parameters are set for each course stage; S130. Dynamic adjustment of enhancement intensity: The enhancement intensity is dynamically increased with the training round t, specifically adjusted according to the following formula: Where A(t) is the dynamic augmentation strength applied in the t-th round of training; A0 is the initial augmentation strength; S140. Progressive Sample Mixing: During the transition between adjacent stages, samples from the previous stage and the next stage are mixed in a preset ratio to guide the model to achieve learning from easy to difficult.

[0011] Preferably, in step S120, the specific rounds of the three course stages are divided as follows: the basic stage is rounds 1 to 15, the intermediate stage is rounds 16 to 35, and the advanced stage is rounds 36 to 50.

[0012] Preferably, in step S200, the specific configuration of the three course stages is as follows: the basic stage uses lightweight geometric augmentation and high-probability hybrid augmentation, combined with mild label smoothing; the intermediate stage uses a random augmentation strategy that includes multiple image operations, and applies low-probability hybrid augmentation; the advanced stage uses adaptive histogram equalization augmentation and mild geometric perturbation, disables hybrid augmentation and label smoothing, and enables a learning rate scheduler to combine cosine annealing learning rate scheduling.

[0013] Preferably, in the basic stage of step S200, the lightweight geometric enhancement includes rotation, translation, and flip operations; the mixed enhancement is Mixup, with a probability set to 0.7; and the label smoothing coefficient is set to 0.05. In the intermediate stage of step S200, the random enhancement strategy is the RandAugment strategy, including at least one of brightness adjustment, contrast perturbation, affine transformation, hue shift, and sharpening. The probability of mixed enhancement is reduced to 0.5. In the high-difficulty stage of step S200, the adaptive histogram equalization enhancement is a contrast-limited adaptive histogram equalization operation, which only acts on the brightness channel of the grayscale image, improving the edge contrast and visual saliency of the lesion area through block-based local equalization.

[0014] Preferably, in step S200, the data augmentation strategy for each course stage is implemented based on a dynamic weight adjustment mechanism, including: S210. Stage weight initialization: Define an augmentation hybrid weight vector W for each course stage i. i : in, These represent the weights for geometric enhancement, color and contrast perturbation, domain adversarial enhancement, and generative enhancement, respectively, with the sum of each weight being 1. The weights for generative enhancement monotonically increase with the training phase. S220. In-phase enhancement execution: Within course phase i, based on its corresponding weight vector W... i Perform corresponding enhancement operations proportionally; S230. Stage performance feedback evaluation: After the training of each course stage is completed, calculate the performance index on the validation set; S240. Dynamic weight update: Based on the performance index obtained in S230, update the enhancement hybrid weights for the next stage using an exponential moving average strategy; The exponential moving average strategy is as follows: Among them, W i W represents the enhanced hybrid weight vector for the current stage; i+1 This represents the updated weight vector for the next stage. For performance feedback weights; This is the smoothing coefficient.

[0015] Preferably, in step S300, a phased training optimization process is adopted within each course stage, including: S310. Initial feature learning stage: using a high learning rate and classification loss, initial training is performed to establish basic classification features; S320. Feature contrast optimization stage: a contrastive loss term is added to the loss function to optimize the objective function to enhance the separability between classes in the feature space; the objective function optimization formula is: Where L is the objective function; L cls For classification loss; L con For feature contrast loss; For feature contrast loss L con The weighting coefficients; S330. Regularization and attention guidance stage: A regularization term is introduced into the loss function, combined with attention mask constraints, to reduce the risk of model overfitting. The specific optimization objective function formula is as follows: Among them, L reg For regularization terms; For regularization term L reg Weighting coefficients; Attention mask constraint Attention constraint The weighting coefficients, and The attention mask constraint increases with the course stage; When there are no segmentation labels, it is defined as: in, The mask is generated by the model's attention module; This refers to the feature maps of the intermediate layers of the model. Masking of lesion areas for annotation or pseudo-annotation; For Hadamard product; These are the attention constraint weighting coefficients, used to balance the relative importance of the non-lesion region suppression term and the lesion region consistency constraint term. The attention mask constraint term When there are lesion mask annotations or pseudo-annotations, it is defined as: in, The Dice loss function measures the difference in spatial overlap between the predicted attention mask and the actual lesion mask. A mask for the true or false annotation of the lesion area; , is the binary cross-entropy loss weight coefficient, used to balance the contributions of Dice loss and cross-entropy loss to the total attention constraint; The binary cross-entropy loss function is used to constrain the classification consistency between the attention mask and the lesion mask pixel by pixel; S340. Parameter transfer and stage switching: When the training of course stage i converges, its model parameters W are... i Used for initialization phase i+1 to achieve continuous knowledge transfer and optimization.

[0016] Preferably, in step S300, dynamically adjusting the enhancement strategy parameters includes dynamically adjusting the probability of hybrid enhancement and the label smoothing coefficient. When hybrid enhancement is used, the specific data mixing operation is implemented according to the following formula: in, The sampling coefficients are in the interval [0,1], and decrease as the course progresses; x 1 and x 2 Two training samples and their labels are provided. The multimodal regularization strategy includes exponential moving average model update, mixed precision training, gradient pruning, and random depth mechanism. The model is a classification model based on the Transformer architecture, with its backbone network based on the Swin-TinyTransformer structure. After removing its pre-trained classification head, initial weights are loaded, and the AdamW optimizer is used for model optimization. The model training uses a cosine annealing learning rate scheduling function, and its learning rate... Updated according to the following formula: in, and , where represents the upper and lower bounds of the learning rate; t represents the current training round; and T represents the total training rounds.

[0017] Preferably, in step S400, the specific steps of generating a heatmap and overlaying it on the original image include: S410. Gradient response calculation: extracting the activation map A of the model convolutional layer for the target category c. k Gradient response: in, The weights are calculated for the gradient; Z is the total number of pixels in the convolutional layer activation map; k represents the index of the convolutional layer activation map; i and j represent the values ​​of the convolutional layer activation map A, respectively. k The specific pixel positions in the height and width dimensions; y represents the predicted score of target category c. c Relative to convolutional layer activation map A k The gradient of the pixel value at position (i,j); S420. Heatmap generation: Calculate weights based on the gradient, perform a weighted summation of the activation features of the convolutional layer activation map, and then process the summation using a linear rectified function to generate a heatmap: in, Represents a heatmap; ReLU represents a linear rectified function; S430. Heatmap overlay visualization: The generated heatmap is upsampled to match the size of the original input image, rendered using color gradient mapping, and then overlaid on the original image; S440. Interpretability quantification analysis: The cross-union ratio (CUI) between the areas marked on the heatmap and the actual labeled lesion areas is calculated, and the response intensity distribution of the heatmap areas is statistically analyzed to generate an interpretable quantification report.

[0018] The technological advancements achieved by this invention are as follows, thanks to the adoption of the above technical solutions.

[0019] This invention significantly improves training stability: through course-based multi-stage augmentation scheduling, it achieves a gradual increase in training difficulty, avoiding the feature confusion caused in the early stage of training and the insufficient data diversity caused in the later stage by traditional fixed augmentation strategies, thus significantly improving the stability and convergence speed of model training.

[0020] This invention enhances generalization ability: the dynamically adjusted enhancement strategy and multimodal regularization technology enable the model to better adapt to the diversity and complexity of medical images, improving the model's generalization ability on medical images acquired under different devices and shooting conditions.

[0021] This invention effectively solves the problem of data leakage: the innovative group-aware hierarchical cross-validation mechanism solves the problem of information leakage between the training set and the validation set for the same patient sample from the source of data partitioning, ensuring the authenticity and reliability of the model evaluation results.

[0022] This invention significantly improves clinical interpretability: through gradient significance heatmap visualization technology, the image regions on which the model's decisions are based can be displayed intuitively, enabling clinicians to understand and verify the basis of the model's judgments, thereby improving the clinical acceptance and credibility of the intelligent diagnostic system.

[0023] This invention has strong adaptive learning capabilities: The dynamic weight update and parameter transfer mechanism introduced in this invention enables the model to adaptively adjust the enhancement strategy according to the training progress and performance feedback, thus achieving a more intelligent and efficient training process. Attached Figure Description

[0024] Figure 1 is a flowchart of the present invention; Figure 2 is a flowchart of step S100 of the present invention, which divides the model training process into three course stages that are executed sequentially and with increasing difficulty according to a predetermined number of rounds; Figure 3 is a flowchart of step S200 of the present invention, which implements the data augmentation strategy for each course stage based on a dynamic weight adjustment mechanism; Figure 4 is a flowchart of step S300 of the present invention, which adopts a phased training optimization process within each course stage; Figure 5 is a flowchart of step S400 of the present invention, which generates a heat map and overlays it on the original image. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0026] A medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation, as shown in Figure 1, begins with the input of the original medical image dataset D, and goes through three stages: augmentation, optimization, and interpretation before outputting the final classification result. .

[0027] The overall process meets the following requirements: in, This represents image data after curriculum-based and hybrid enhancement processing; These are the optimal parameters obtained through multi-stage training. Specifically, the following steps are included: S100. Construction of a curriculum-based data augmentation mechanism: Load the medical image dataset, and use group-aware hierarchical cross-validation to divide the dataset into a training set and a validation set, ensuring that images of the same patient do not appear in the training set and the validation set at the same time; divide the model training process into three curriculum stages that are executed sequentially and with increasing difficulty according to a predetermined number of rounds.

[0028] In this step, group-aware stratified cross-validation is achieved as follows: extracting the stem value of the medical image sample file name as the patient identifier; grouping the samples based on the stem value so that images of the same patient belong to the same group; and maintaining the integrity of samples within the group when dividing the training set and validation set to ensure that samples from the same group appear only in either the training set or the validation set.

[0029] As shown in Figure 2, the model training process is divided into three course stages that are executed sequentially and with increasing difficulty according to a predetermined number of rounds. Specifically, it includes: S110. Sample complexity assessment: Based on the feature entropy, edge clarity and lesion salience of the image, the training samples are assessed for difficulty and stratified.

[0030] Specifically, based on a dynamic hierarchical strategy of sample complexity, the training samples are divided into three sets according to feature difficulty: low difficulty, medium difficulty, and high difficulty, denoted as follows: in, This represents a subset of low-difficulty samples, including image samples with clear boundaries and significant lesions; This represents a subset of high-difficulty samples, including samples with high noise or blurred lesions; This represents a subset of samples of moderate difficulty, containing image samples of moderate difficulty.

[0031] S120. Stage Division and Strategy Configuration: Based on the difficulty assessment results, the model training process is divided into a basic stage, an intermediate stage, and a high-difficulty stage, and enhancement types and parameters are set for each stage.

[0032] Specifically, the three course stages are divided into rounds as follows: the basic stage consists of rounds 1 to 15, the intermediate stage consists of rounds 16 to 35, and the advanced stage consists of rounds 36 to 50.

[0033] S130. Dynamic Adjustment of Enhancement Intensity: The enhancement intensity is dynamically increased with the training round t, specifically adjusted according to the following formula: Where A(t) is the dynamic augmentation strength applied in the t-th round of training; A0 is the initial augmentation strength; This is the course growth coefficient; this mechanism enables a progressive learning process for the model from "easy to difficult," thereby improving training stability and feature abstraction capabilities.

[0034] S140. Progressive Sample Mixing: During the transition between adjacent stages, samples from the previous stage and the next stage are mixed in a preset ratio to guide the model to learn from easy to difficult.

[0035] S200. Implementation of phased hybrid augmentation strategy: Apply data augmentation strategies with different intensities and data distributions at different course stages, and the complexity and regularization strength of the augmentation strategies change progressively with each stage.

[0036] In this step, the specific configuration for the three course stages is as follows: The basic stage uses lightweight geometric enhancement and high-probability mixed enhancement, combined with mild label smoothing. Specifically, lightweight geometric enhancement includes rotation, translation, and flip operations; the mixed enhancement is Mixup, with a probability set to 0.7; and the label smoothing coefficient is set to 0.05.

[0037] In the intermediate stage, a random augmentation strategy incorporating multiple image operations is used, along with a low-probability mixed augmentation. Specifically, the random augmentation strategy is the RandAugment strategy, which includes at least one of brightness adjustment, contrast perturbation, affine transformation, hue shift, and sharpening, to improve the model's adaptability to changes in imaging conditions; the mixed augmentation is called Mixup, with the probability reduced to 0.5.

[0038] In the high-difficulty stage, adaptive histogram equalization enhancement with mild geometric perturbation is used, blending enhancement and label smoothing are turned off, and a learning rate scheduler is enabled to combine cosine annealing learning rate scheduling. Specifically, adaptive histogram equalization enhancement is a contrast-limited adaptive histogram equalization operation (CLAHE), which only operates on the luminance channel of the grayscale image, improving the edge contrast and visual saliency of the lesion area through block-based local equalization.

[0039] As shown in Figure 3, the data augmentation strategies for each course stage are implemented based on a dynamic weight adjustment mechanism, including: S210. Stage weight initialization: Define an augmentation hybrid weight vector W for each course stage i. i : in, These represent the weights for geometric enhancement (rotation, scaling), color and contrast perturbation, domain adversarial enhancement, and generative enhancement (GAN-generated samples), respectively; and the sum of all weights is 1. The weights for generative enhancement monotonically increase with the training phase, i.e., satisfying: in, This represents the sum of all weights; This represents the weight coefficient of generative augmentation (such as GAN-generated samples) in the i-th training phase; This represents the weight coefficients of generative augmentation in the (i+1)th training phase (i.e., the next phase).

[0040] This enables dynamic transfer from basic structure learning to domain generalization.

[0041] S220. Enhanced Execution within a Stage: Within course stage i, based on its corresponding weight vector W i The corresponding enhancement operations are performed proportionally, including rotation, contrast changes, artifact simulation, and GAN-generated samples.

[0042] S230. Stage Performance Feedback Evaluation: After the training of each course stage is completed, calculate the performance metrics on the validation set.

[0043] Specifically, the accuracy of the validation set and the loss fluctuation are calculated.

[0044] S240. Dynamic Weight Update: Based on the performance metrics obtained in S230, the enhanced hybrid weights for the next stage are updated using an exponential moving average strategy; the exponential moving average strategy is as follows:

[0045] Among them, W i W represents the enhanced hybrid weight vector for the current stage; i+1 This represents the updated weight vector for the next stage. For performance feedback weights; The smoothing coefficient is (0.7–0.9).

[0046] Through the above steps, a self-learning course scheduling mechanism is formed, which can adaptively adjust the enhancement type and its proportion at each stage based on the performance of the validation set.

[0047] S300. Multi-stage model training and optimization: In each of the divided course stages, the enhancement strategy parameters are dynamically adjusted, and a multimodal regularization strategy is used to train and optimize the backbone network of the model to improve the model's stability and robustness.

[0048] In this step, dynamically adjusting the enhancement strategy parameters includes dynamically adjusting the probability of hybrid enhancement and the label smoothing coefficient. When using hybrid enhancement, the specific data mixing operation is implemented according to the following formula: in, The sampling coefficients are in the interval [0,1] and decrease as the course progresses to control the degree of sample mixing and label fusion intensity; x 1 and x 2 These are two training samples and their labels.

[0049] Multimodal regularization strategies include exponential moving average (EMA) model updates, mixed precision training (AMP), gradient pruning, and stochastic depth mechanisms.

[0050] The model is a classification model based on the Transformer architecture. Its backbone network is based on the Swin-TinyTransformer structure, with its pre-trained classification heads removed and initial weights loaded. The AdamW optimizer is used for model optimization. The model training employs a cosine annealing learning rate scheduling function to smoothly reduce the learning rate and enhance the model's convergence stability. Updated according to the following formula: in, and , where represents the upper and lower bounds of the learning rate; t represents the current training round; and T represents the total training rounds.

[0051] As shown in Figure 4, a phased training optimization process is adopted within each course stage to improve feature diversity and significantly enhance the model's cross-domain generalization performance. Specifically, it includes: S310. Initial feature learning stage: Using a high learning rate and classification loss, initial training is performed to establish basic classification features.

[0052] S320. Feature Contrast Optimization Stage: A contrastive loss term is added to the loss function to optimize the objective function and enhance the separability between classes in the feature space; the objective function optimization formula is: Where L is the objective function; L cls For classification loss; L con For feature contrast loss; For feature contrast loss L con The weighting coefficients.

[0053] S330. Regularization and Attention Guidance Stage: A regularization term is introduced into the loss function, combined with attention mask constraints, to reduce the risk of model overfitting. The specific optimization objective function formula is as follows: Among them, L reg For regularization terms; For regularization term L reg Weighting coefficients; This is an attention mask constraint term used to guide the model to enhance representations in lesion-related regions and suppress irrelevant responses in background regions. Attention constraint The weighting coefficients, and As the course stages increase, it remains consistent with the course-based enhanced scheduling (S100, S200).

[0054] Attention mask constraint In scenarios without segmentation annotation, the following formula is preferred: in, The mask is generated by the model's attention module; This refers to the feature maps of the intermediate layers of the model. A mask for labeled or pseudo-labeled lesion regions (or target features obtained by lesion regions); For Hadamard product; These are the attention constraint weighting coefficients, used to balance the relative importance of the non-lesion region suppression term and the lesion region consistency constraint term. Attention mask constraint When there are lesion mask annotations or pseudo-annotations, it is defined as: in, The Dice loss function measures the difference in spatial overlap between the predicted attention mask and the actual lesion mask. A mask for the true or false annotation of the lesion area; , is the binary cross-entropy loss weight coefficient, used to balance the contributions of Dice loss and cross-entropy loss to the total attention constraint; The binary cross-entropy loss function is used to constrain the classification consistency between the attention mask and the lesion mask on a pixel-by-pixel basis.

[0055] S340. Parameter Transfer and Stage Switching: After the training of course stage i converges, its model parameters W are transferred. i Used for initialization phase i+1, forming the progressive transfer training structure: This maintains the continuity of the feature space and the stability of optimization convergence, enabling continuous knowledge transfer and optimization.

[0056] S400. Model Interpretability Output Generation: The trained model is used to classify medical images and generate heatmaps based on the gradient information output by the model. These heatmaps are then overlaid on the original images to visualize the lesion areas on which the model makes decisions, thereby assisting doctors in judging the distribution of lesions and improving medical interpretability.

[0057] As shown in Figure 5, the specific steps for generating a heatmap and overlaying it onto the original image include: S410. Gradient response calculation: Extracting the activation map A of the model convolutional layer for the target category c. k Gradient response: in, The weights are calculated for the gradient; Z is the total number of pixels in the convolutional layer activation map; k represents the index of the convolutional layer activation map; i and j represent the values ​​of the convolutional layer activation map A, respectively. k The specific pixel positions in the height and width dimensions; y represents the predicted score of target category c. cRelative to convolutional layer activation map A k The gradient of the pixel value at position (i,j).

[0058] S420. Heatmap Generation: Based on gradient calculation of weights, the activation features of the convolutional layer activation map are weighted and summed, and then processed by a linear rectified function to generate a heatmap. in, Represents a heatmap; ReLU represents the linear rectified function, used to preserve the region of positive contribution.

[0059] S430. Heatmap Overlay Visualization: The generated heatmap is upsampled to match the size of the original input image, and then rendered using color gradient mapping to highlight the lesion area. It is then overlaid on the original image to achieve an intuitive visualization output of the model's area of ​​interest.

[0060] S440. Quantitative Analysis of Interpretability: Calculate the Intersection over Union (IoU) between the area marked on the heatmap and the actual marked lesion area, and statistically analyze the response intensity distribution of the heatmap area to generate an interpretable quantitative report.

[0061] A medical image classification system based on curriculum-based augmentation scheduling and hybrid augmentation includes an input module, a curriculum-based augmentation scheduling module, a hybrid augmentation control module, a model training module, and an interpretable visualization module. These modules sequentially execute steps S100 to S400 to achieve curriculum-based augmentation-driven intelligent classification of medical images.

Claims

1. A medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation, characterized in that: Includes the following steps: S100. Construction of Curriculum-Based Data Augmentation Mechanism: Load the medical image dataset and use group-aware hierarchical cross-validation to divide the dataset into training and validation sets, ensuring that images of the same patient do not appear in the training and validation sets simultaneously; divide the model training process into three sequentially executed, progressively more difficult course stages according to predetermined rounds; S200. Implementation of Phased Hybrid Augmentation Strategy: Apply data augmentation strategies with different intensities and data distributions in different course stages, with the complexity and regularization strength of the augmentation strategies varying progressively with each stage; S300. Multi-Stage Model Training and Optimization: Dynamically adjust the augmentation strategy parameters in each of the divided course stages, and use a multimodal regularization strategy to train and optimize the model's backbone network; S400. Generation of Interpretable Model Output: Use the trained model to classify medical images and generate heatmaps based on the gradient information output by the model, overlaying them onto the original images to visualize the lesion regions on which the model's decisions are based.

2. The medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation according to claim 1, characterized in that: In step S100, group-aware hierarchical cross-validation is achieved in the following way: extracting the stem value of the medical image sample file name as the patient identifier; grouping the samples based on the stem value so that the images of the same patient belong to the same group; and maintaining the integrity of the samples within the group when dividing the training set and the validation set to ensure that the same group of samples appears only in the training set or the validation set.

3. The medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation according to claim 1, characterized in that: In step S100, the model training process is divided into three sequentially executed, progressively more difficult stages according to a predetermined number of rounds. Specifically, this includes: S110. Sample complexity assessment: Based on image feature entropy, edge sharpness, and lesion salience indicators, the training samples are assessed for difficulty and stratified; S120. Stage division and strategy configuration: Based on the difficulty assessment results, the model training process is divided into a basic stage, a medium stage, and a high-difficulty stage, and enhancement types and parameters are set for each stage; S130. Dynamic adjustment of enhancement intensity: The enhancement intensity is dynamically increased with each training round t, specifically adjusted according to the following formula: Where A(t) is the dynamic augmentation strength applied in the t-th round of training; A0 is the initial augmentation strength; S140. Progressive Sample Mixing: During the transition between adjacent stages, samples from the previous stage and the next stage are mixed in a preset ratio to guide the model to achieve learning from easy to difficult.

4. The medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation according to claim 3, characterized in that: In step S120, the specific rounds of the three course stages are divided as follows: the basic stage consists of rounds 1 to 15, the intermediate stage consists of rounds 16 to 35, and the advanced stage consists of rounds 36 to 50.

5. The medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation according to claim 3, characterized in that: In step S200, the specific configuration of the three course stages is as follows: In the basic stage, lightweight geometric enhancement and high-probability hybrid enhancement are used, combined with mild label smoothing; in the intermediate stage, a random enhancement strategy containing multiple image operations is used, and low-probability hybrid enhancement is applied. In the high-difficulty phase, adaptive histogram equalization enhancement with mild geometric perturbation is used, hybrid enhancement and label smoothing are turned off, and a learning rate scheduler is enabled to combine cosine annealing learning rate scheduling.

6. The medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation according to claim 5, characterized in that: In the basic stage of step S200, lightweight geometric enhancement includes rotation, translation, and flipping operations; mixed enhancement is Mixup with a probability set to 0.7; and the coefficient for label smoothing is set to 0.

05. In the intermediate stage of step S200, the random enhancement strategy is the RandAugment strategy, which includes at least one of brightness adjustment, contrast perturbation, affine transformation, hue shift, and sharpening; and the probability of mixed enhancement is reduced to 0.

5. In the advanced stage of step S200, adaptive histogram equalization enhancement is a contrast-limited adaptive histogram equalization operation that only acts on the brightness channel of the grayscale image, improving the edge contrast and visual saliency of the lesion area through block-based local equalization.

7. A medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation as described in claim 6, characterized in that: In step S200, the data augmentation strategy for each course stage is implemented based on a dynamic weight adjustment mechanism, including: S210. Stage weight initialization: Define an augmentation hybrid weight vector W for each course stage i. i : in, These represent the weights for geometric enhancement, color and contrast perturbation, domain adversarial enhancement, and generative enhancement, respectively, with the sum of each weight being 1. The weights for generative enhancement monotonically increase with the training phase. S220. In-phase enhancement execution: Within course phase i, based on its corresponding weight vector W... i Perform corresponding enhancement operations proportionally; S230. Stage performance feedback evaluation: After the training of each course stage is completed, calculate the performance index on the validation set; S240. Dynamic weight update: Based on the performance index obtained in S230, update the enhancement hybrid weights for the next stage using an exponential moving average strategy; The exponential moving average strategy is as follows: Among them, W i W represents the enhanced hybrid weight vector for the current stage; i+1 This represents the updated weight vector for the next stage. For performance feedback weights; This is the smoothing coefficient.

8. A medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation according to claim 7, characterized in that: In step S300, a phased training optimization process is adopted within each course stage, including: S310. Initial feature learning stage: using a high learning rate and classification loss, initial training is performed to establish basic classification features; S320. Feature contrast optimization stage: a contrastive loss term is added to the loss function to optimize the objective function to enhance the separability between classes in the feature space; the optimization formula for the objective function is: Where L is the objective function; L cls For classification loss; L con For feature contrast loss; Feature contrast loss L con The weighting coefficients; S330. Regularization and attention guidance stage: A regularization term is introduced into the loss function, combined with attention mask constraints, to reduce the risk of model overfitting. The specific optimization objective function formula is as follows: Among them, L reg For regularization terms; For regularization term L reg Weighting coefficients; For attention mask constraints; Attention constraint The weighting coefficients, and The attention mask constraint increases with the course stage; When there are no segmentation labels, it is defined as: in, The mask is generated by the model's attention module; This refers to the feature maps of the intermediate layers of the model. Masking of lesion areas for annotation or pseudo-annotation; For Hadamard product; These are the attention constraint weighting coefficients, used to balance the relative importance of the non-lesion region suppression term and the lesion region consistency constraint term. The attention mask constraint term When there are lesion mask annotations or pseudo-annotations, it is defined as: in, The Dice loss function measures the difference in spatial overlap between the predicted attention mask and the actual lesion mask. A mask for the true or false annotation of the lesion area; , is the binary cross-entropy loss weight coefficient, used to balance the contributions of Dice loss and cross-entropy loss to the total attention constraint; The binary cross-entropy loss function is used to constrain the classification consistency between the attention mask and the lesion mask pixel by pixel; S340. Parameter transfer and stage switching: When the training of course stage i converges, its model parameters W are... i Used for initialization phase i+1 to achieve continuous knowledge transfer and optimization.

9. A medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation as described in claim 8, characterized in that: In step S300, dynamically adjusting the enhancement strategy parameters includes dynamically adjusting the probability of hybrid enhancement and the label smoothing coefficient. When using hybrid enhancement, the specific data mixing operation is implemented according to the following formula: in, The sampling coefficients are in the interval [0,1], and decrease as the course progresses; x 1 and x 2 Two training samples and their labels are provided. The multimodal regularization strategy includes exponential moving average model update, mixed precision training, gradient pruning, and random depth mechanism. The model is a classification model based on the Transformer architecture, with its backbone network based on the Swin-TinyTransformer structure. After removing its pre-trained classification head, initial weights are loaded, and the AdamW optimizer is used for model optimization. The model training uses a cosine annealing learning rate scheduling function, and its learning rate... Updated according to the following formula: in, and , where represents the upper and lower bounds of the learning rate; t represents the current training round; and T represents the total training rounds.

10. A medical image classification method based on curriculum-based augmentation scheduling and hybrid augmentation according to claim 1, characterized in that: In step S400, the specific steps for generating a heatmap and overlaying it onto the original image include: S410. Gradient response calculation: Extracting the activation map A of the model convolutional layer for target category c. k Gradient response: in, The weights are calculated for the gradient; Z is the total number of pixels in the convolutional layer activation map; k represents the index of the convolutional layer activation map; i and j represent the values ​​of the convolutional layer activation map A, respectively. k The specific pixel positions in the height and width dimensions; y represents the predicted score of target category c. c Relative to convolutional layer activation map A k The gradient of the pixel value at position (i,j); S420. Heatmap generation: Calculate weights based on the gradient, perform a weighted summation of the activation features of the convolutional layer activation map, and then process the summation using a linear rectified function to generate a heatmap: in, Represents a heatmap; ReLU represents a linear rectified function; S430. Heatmap overlay visualization: The generated heatmap is upsampled to match the size of the original input image, rendered using color gradient mapping, and then overlaid on the original image; S440. Interpretability quantification analysis: The cross-union ratio (CUI) between the areas marked on the heatmap and the actual labeled lesion areas is calculated, and the response intensity distribution of the heatmap areas is statistically analyzed to generate an interpretable quantification report.

Images