Intelligent diagnosis method and system for cervical OCT (Optical Coherence Tomography) image

Abstract

The invention discloses an intelligent diagnosis method and system for a cervical OCT (Optical Coherence Tomography) image. Comprising the following steps: (1) constructing a deep evidence learning network architecture; (2) designing a double knowledge distillation mechanism; (3) designing an adaptive dynamic weight adjustment strategy; (4) introducing interpretable loss enhanced feature learning; and evaluating the generalization ability and the anomaly detection performance of the model. The technical problems that an existing medical image classification model is insufficient in uncertainty quantization, single in knowledge distillation mechanism and insufficient in interpretability are solved, and a deployable technical scheme is provided for medical image intelligent diagnosis under privacy protection.

Description

Technical Field

[0001] This invention relates to the fields of medical image analysis and artificial intelligence technology, and in particular to an intelligent diagnostic method for cervical optical coherence tomography (OCT) images based on deep evidence learning and interpretable knowledge distillation. This method is suitable for achieving accurate screening and classification of cervical lesions while protecting patient privacy, and provides prediction uncertainty estimation and interpretability analysis. Background Technology

[0002] Traditional cervical cancer screening methods, such as HPV testing, liquid-based thin-layer cytology (TCT), and colposcopy, all have limitations, including invasive procedures, reliance on human experience for diagnosis, and high false-positive rates. In recent years, optical coherence tomography (OCT) has shown significant advantages in imaging the microstructure of cervical tissue due to its micron-level resolution and non-invasive nature. However, its clinical application is still limited by doctors' insufficient understanding of OCT image features and the low efficiency of manual image interpretation.

[0003] Currently, deep learning-based medical image classification technology has made groundbreaking progress in image diagnosis in fields such as ophthalmology and pulmonary medicine. However, intelligent analysis of cervical OCT images still faces multiple challenges: on the one hand, medical image data annotation is costly and the amount of data is limited, which can easily lead to model overfitting; on the other hand, traditional deep learning models lack the ability to quantify uncertainty, making it difficult to assess prediction confidence and limiting the reliability of clinical applications; in addition, existing models are mostly black-box structures, lacking interpretability and failing to meet the clinical needs for diagnostic evidence.

[0004] Deep Evidential Learning (DEL) offers a novel approach to addressing the aforementioned problems by modeling classification uncertainty through the del distribution. Knowledge distillation, a technique that utilizes a teacher-student network architecture, achieves model compression and performance enhancement. However, existing methods still suffer from the following drawbacks: 1) A simplistic knowledge distillation mechanism: Most methods employ only probabilistic knowledge distillation, neglecting the knowledge transfer of evidence vectors and failing to fully leverage the advantages of deep evidence learning models; 2) Rigid weight adjustment: Traditional methods use fixed weights to combine knowledge distillation loss and hard classification loss, making it impossible to adaptively adjust based on training progress; 3) Insufficient interpretability: Existing methods lack interpretable modeling of the intrinsic structure of features and class differences, making it difficult to meet the clinical needs for diagnostic evidence.

[0005] Therefore, there is an urgent need for a deep learning method that balances uncertainty quantification, knowledge distillation optimization, and interpretability enhancement to promote the clinical application of cervical OCT intelligent diagnostic technology. Summary of the Invention

[0006] Through extensive research and practice, the inventors of this application have discovered that existing cervical OCT image classification methods mainly rely on single probabilistic knowledge distillation and are limited by fixed weight adjustment strategies. This results in the model being unable to adaptively balance teacher model guidance and real label learning during training, and lacks interpretable modeling of the intrinsic structure of features and class differences, making it difficult to meet the multiple clinical needs for diagnostic accuracy, uncertainty quantification, and interpretability.

[0007] Therefore, to address the above problems, this invention proposes a cervical OCT-assisted diagnostic method based on deep evidence learning and interpretable knowledge distillation, which effectively improves the model's diagnostic accuracy, uncertainty quantification ability, and interpretability of cervical OCT images, providing a feasible path for the construction of an intelligent screening system for cervical lesions.

[0008] The overall inventive concept of this invention is as follows: 1) Construct a deep evidence learning network architecture, using ResNet18 as the image encoder to extract features, outputting logists through a fully connected network, and outputting evidence vectors through the Softplus activation function, modeling classification uncertainty based on del distribution; 2) Design a dual knowledge distillation mechanism, introducing probability knowledge distillation (Probability KD) and evidence knowledge distillation (Evidence KD) for collaborative optimization, combining the two distillation losses through independent weights or complementary modes to achieve model compression and performance improvement; 3) Use a cosine annealing mechanism to dynamically adjust the knowledge distillation weights according to the number of training steps, achieving an adaptive balance between emphasizing the soft target of the teacher model in the early stage of training and emphasizing the hard target of the real label in the later stage of training; 4) Introduce interpretability loss to enhance feature learning, learning the channel relationship and spatial saliency distribution of feature maps through intra-instance latent feature knowledge (ILFK) loss, and modeling the differences in category feature distribution through inter-class variation fuzzy perception (IVFP) loss; 5) Use a cosine scheduler to adaptively adjust the learning rate and weight decay, automatically optimizing the training process according to the training progress.

[0009] To address the aforementioned technical problems, this invention provides, in one aspect, an intelligent diagnostic method for cervical OCT images based on deep evidence learning and interpretable knowledge distillation, comprising: Acquire cervical OCT images and perform data augmentation; A deep evidence learning network architecture is constructed, including an image encoder and an evidence classifier. The image encoder is used to extract features from cervical OCT images, and the evidence classifier converts the features into raw output values ​​logits and evidence vectors, and models the classification uncertainty based on the Dirichlet distribution. A dual knowledge distillation training framework is designed, which introduces probabilistic knowledge distillation and evidence knowledge distillation for collaborative optimization, thereby achieving network architecture compression and performance improvement. Both the teacher network and the student network are based on a deep evidence learning architecture. An adaptive dynamic weight adjustment strategy is designed, which uses a cosine annealing mechanism to dynamically adjust the knowledge distillation weights according to the number of training steps, balancing the knowledge distillation loss and the hard classification loss, thereby obtaining the base loss. Interpretability loss is introduced to enhance feature learning, including intra-instance latent feature knowledge loss and inter-class variation fuzzy perception loss, and combined with the base loss to construct the final total loss; The cosine scheduler is designed to adaptively adjust the learning rate and weight decay, and automatically optimize the overall training process according to the training progress. The trained deep evidence learning model is used to classify and predict the OCT images to be tested, and the classification results and uncertainty estimates are output.

[0010] Furthermore, the image encoder uses ResNet18 to extract features from cervical OCT images, and the evidence classifier converts the features into logits through a multi-layer fully connected network. The logits are then used to output an evidence vector through the Softplus activation function. The evidence vector is converted into the parameters α of the Dirichlet distribution by adding 1, thus realizing uncertainty modeling.

[0011] Furthermore, the formula for calculating probabilistic knowledge distillation is as follows:

[0012] Where T is the temperature parameter for probabilistic knowledge distillation, used to soften the probability distribution; s represents the logits output by the student network, t represents the logits output by the teacher network, and KL(· ·) represents the KL divergence, which measures the difference between two probability distributions; For activation functions; The formula for calculating evidence knowledge distillation is:

[0013] in Temperature parameters for distillation of evidence knowledge; This represents the evidence vector output by the student network. KL(·) represents the evidence vector output by the teacher network. ·) represents the KL divergence, which measures the difference between two probability distributions; For activation functions; Knowledge distillation loss is obtained by combining probabilistic knowledge distillation and evidential knowledge distillation using independent weighting patterns or complementary patterns.

[0014] Furthermore, the formula for calculating the knowledge distillation weight is as follows:

[0015]

[0016] in This represents the current number of training steps. Total training steps Indicates the weighting adjustment factor. This represents the minimum weight for knowledge distillation. This represents the maximum weight for knowledge distillation. Balancing the knowledge distillation loss and hard classification loss is achieved through the following formula:

[0017] in, Basic loss, hard classification loss The loss is cross-entropy, and w is the dynamically adjusted knowledge distillation weight. This is due to knowledge distillation loss.

[0018] Furthermore, in-instance latent feature knowledge loss Loss due to channel relationship Loss of spatial relationship The system comprises three components: a feature map for learning the intrinsic structural patterns of the encoder output; a channel relation loss for learning channel dependencies by calculating the cosine similarity matrix between channels of the feature map; and a spatial relation loss for learning the spatial saliency distribution by calculating the average value of the feature map along the channel dimension. The specific calculation formula is as follows:

[0019] in:

[0020]

[0021] Where K is the number of channels, and H and W are the feature map space dimensions; This represents the channel relationship matrix of the b-th sample in the teacher network. Let B represent the channel relationship matrix of the b-th sample in the student network, where B represents the number of samples. The spatial saliency map of a single channel is obtained by averaging the K channel feature maps of the b-th sample in the teacher network along the channel dimension. The spatial saliency map is obtained by averaging the K channel feature maps of the b-th sample in the student network along the channel dimension.

[0022] Furthermore, fuzzy perception of inter-class changes The loss is modeled using style subsets to represent the categorical feature distribution. Style centers and styles are defined, and style subset parameters are set using the `set_style_subsets` method, supporting dynamic updates. These are periodically updated based on the feature distribution of the teacher network, as shown in the formula: For positive samples:

[0023] Where C is the number of categories in the classification task; For negative samples:

[0024] in This represents the distance between the i-th sample feature and the j-th style center. Let be the j-th style radius, and α and β be hyperparameters. The distance between the sample and the nearest style center. The radius of the nearest style center; .

[0025] Furthermore, the final total loss is:

[0026] in Based on the loss, and These are the in-instance latent feature knowledge loss and the inter-class change fuzzy perception loss, respectively. and These are the weighting coefficients.

[0027] Furthermore, the learning rate is set as follows:

[0028] Where t is the current training step count, and T is the total training steps. Based on the learning rate, Minimum learning rate; A warm-up mechanism is used during training, and the same cosine scheduling strategy as the learning rate is adopted. The weight decay coefficient is automatically adjusted according to the training progress, and the weight decays from the initial value to the final value according to the cosine function.

[0029] The present invention also provides a cervical OCT image intelligent diagnostic system, comprising a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute a cervical OCT image intelligent diagnostic method as described in the above technical solution.

[0030] The present invention also provides a computer-readable storage medium, including a readable storage medium on which a computer program is stored, wherein when the computer program is executed, it implements a cervical OCT image intelligent diagnosis method as described in the above technical solution.

[0031] The cervical tissue OCT image classification method provided by this invention significantly improves the diagnostic accuracy, uncertainty quantification ability, and interpretability of the model for cervical OCT images within a deep evidence learning framework through a dual knowledge distillation, adaptive weight adjustment, and interpretability loss enhancement mechanism. This method innovatively constructs a dual distillation mechanism that collaboratively optimizes probabilistic knowledge distillation and evidential knowledge distillation. By combining the two distillation losses with independent weights or complementary modes, it fully utilizes the advantages of deep evidence learning models. A cosine annealing mechanism dynamically adjusts the knowledge distillation weights based on the number of training steps, achieving adaptive balance of losses during training and enabling stable optimization of the training process without manual parameter tuning. Interpretive loss is introduced to enhance feature learning. Intra-instance latent feature knowledge (ILFK) loss is used to learn the channel relationships and spatial saliency distribution of feature maps, while inter-class variation fuzzy perception (IVFP) loss is used to model the differences in category feature distribution, improving the model's interpretability of the intrinsic structure of features and category differences. Through multi-fold cross-validation and evaluation on multiple datasets, this method significantly improves upon traditional methods in core metrics such as accuracy and AUC. It solves the technical challenges of insufficient uncertainty quantification, a single knowledge distillation mechanism, and insufficient interpretability in existing medical image classification models, providing a deployable technical solution for intelligent diagnosis of medical images under privacy protection. Attached Figure Description

[0032] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0033] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0034] The method proposed in this invention constructs a multi-layered, adaptive knowledge transfer and feature learning framework by organically integrating Probability KD, Evidence KD, adaptive dynamic weight adjustment, and interpretability loss (ILFK and IVFP). The mechanism by which the components work together is as follows: I. Synergistic Mechanism of Dual Knowledge Distillation 1. The complementary relationship between probabilistic knowledge distillation and evidential knowledge distillation Probability knowledge distillation (KD) and evidence knowledge distillation (KD) form a complementary knowledge transfer mechanism within the deep evidence learning framework: 1) Probabilistic KD level: The soft label probability distributions of the teacher network and student network are aligned through KL divergence to convey knowledge at the "classification decision" level. Soft labels provide richer information on inter-class relationships through temperature scaling, enabling the student network to learn the teacher network's "soft judgment" on samples.

[0035] 2) Evidence KD Level: By aligning the evidence vectors of the teacher and student networks using KL divergence or mean squared error, knowledge at the "uncertainty quantification" level is conveyed. The evidence vector directly reflects the model's confidence in the prediction, enabling the student network to not only learn the classification results but also how to assess the uncertainty of the prediction.

[0036] 3) Synergistic effect: The combination of the two distillation losses is achieved through an independent weighting mode or a complementary mode:

[0037] in For probability KD weights, The KD weights are used as evidence. This design enables student networks to simultaneously learn the classification decision-making patterns and uncertainty assessment capabilities of teacher networks, achieving more comprehensive knowledge transfer.

[0038] II. The Coordinating Role of Adaptive Dynamic Weight Adjustment 2. Adaptive equilibrium of cosine annealing mechanism for double distillation Adaptive dynamic weight adjustment uses a cosine annealing mechanism to dynamically adjust the weight ratio of knowledge distillation loss to hard classification loss based on training progress. (1) In the early stage of training (current_step = 0): the weight w reaches its maximum value. (0.8) The model mainly relies on the soft objective (probability KD + evidence KD) of the teacher network for learning, and quickly obtains the knowledge of the teacher network.

[0039] (2) Mid-training phase (current_step = total_steps / 2): Weight w decreases to its minimum value (0.2) The model mainly relies on hard targets of real labels for learning, which enhances the ability to fit real data.

[0040] (3) Late training phase (current_step = total_steps): The weights w increase again to The model is further optimized based on existing knowledge to achieve fine-tuning.

[0041] Weight calculation formula:

[0042]

[0043] Basic loss:

[0044] This adaptive mechanism ensures a dynamic balance between knowledge distillation and real label learning during training, avoiding training instability caused by fixed weights.

[0045] III. Feature Enhancement Mechanisms for Explainability Loss 3. Multi-dimensional feature learning of ILFK loss and IVFP loss Interpretability loss enhances feature learning from different dimensions, complementing knowledge distillation: (1) ILFK loss (in-instance latent feature knowledge): learns the intrinsic structural patterns of feature maps through channel relation loss (L_CR) and spatial relation loss (L_SR). ILFK loss enables the student network to learn not only the output of the teacher network, but also the channel dependencies and spatial saliency distribution of the teacher network's feature maps, achieving a deeper level of feature alignment.

[0046] (2) IVFP loss (inter-class variation fuzzy perception): It models the differences in the distribution of class features by using style subsets, and adopts a rejection strategy (away from the style center) for positive samples and an attraction strategy (closer to the style center) for negative samples. IVFP loss enhances the model's learning of class boundaries and improves the discriminativeness of classification.

[0047] (3) Synergy with knowledge distillation: ILFK loss acts on the feature level, complementing the knowledge transfer at the output level of probability KD and evidence KD, thus achieving comprehensive knowledge transfer from features to output.

[0048] IV. Comprehensive Expression of the Overall Loss Function The final total loss function integrates all components, forming a unified multi-level optimization objective:

[0049] in: First layer: For adaptive weighted double knowledge distillation loss; Second layer: The adaptive weighted hard classification loss; Third layer: For interpretability loss (feature level); Fourth layer: For interpretability loss (category level); V. Dynamic Collaborative Process During Training During training, the components work together according to the following process: 1. Forward propagation phase: The student network processes the input image and outputs logits and evidence vectors. The teacher network (if provided) processes the same input and outputs logits and evidence vectors. Extract feature maps from the student and teacher networks (for ILFK loss). 2. Loss Calculation Stage: Calculate probabilistic KD loss: Align soft label distribution Calculate the KD loss of evidence: Align the evidence vector Combined double distillation losses:

[0050] Calculate the hard classification loss:

[0051] Calculate adaptive weights w based on the current training steps. Calculate the base loss:

[0052] Calculating ILFK loss: Learning the intrinsic structure of feature maps Calculate IVFP loss: Learn the distribution of class features Total portfolio loss:

[0053] 3. Backpropagation phase: 1) Backpropagate the total loss to update student network parameters; 2) The adaptive learning rate scheduler adjusts the learning rate according to the training progress; 4. Dynamic weight adjustment: (1) The weight w is updated according to the current step number in each training step; (2) The learning rate is automatically adjusted based on the cosine scheduler; VI. Advantages of the Collaborative Mechanism Through the above-mentioned collaborative mechanism, the present invention achieves the following advantages: 1. Multi-level knowledge transfer: From the output level (probability KD + evidence KD) to the feature level (ILFK), a comprehensive knowledge transfer is achieved.

[0054] 2. Adaptive training strategy: Through cosine annealing, the training process automatically balances teacher guidance and real label learning, without the need for manual parameter tuning.

[0055] 3. Enhanced interpretability: ILFK and IVFP losses not only improve model performance but also enhance model interpretability, meeting clinical needs.

[0056] 4. Uncertainty Quantification: Evidence KD enables students to assess uncertainty in online learning and supports predictive confidence output.

[0057] 5. Training stability: Adaptive weight adjustment avoids training instability caused by fixed weights, and the cosine scheduler ensures smooth changes in the learning rate.

[0058] VII. Parameter Configuration in Practical Applications In practical applications, the typical configuration of each component parameter is as follows: Probability KD parameters: Temperature T = 3.0, weights p w = 0.1 (independent mode) or automatically calculated (complementary mode); Evidence KD parameter: temperature T e = 3.0 (KL mode) or direct MSE, weight e w = 0.1; Adaptive weight parameters: w min = 0.2, w max = 0.8, enable cosine annealing; Interpretability loss parameters: λ ILFK = 100.0, λ IVFP = 1.0 (adjusted according to the task); Learning rate scheduling: base learning rate lr = 10 -4 Minimum learning rate lr min = 10 -6 The number of warmup epochs is 5. Through the above-mentioned comprehensive collaborative mechanism, this invention achieves the organic integration of probabilistic knowledge distillation, evidence knowledge distillation, adaptive weight adjustment and interpretability loss, forming an efficient, stable and interpretable deep evidence learning training framework.

[0059] Specifically, Figure 1 The overall flowchart of a cervical OCT image intelligent diagnosis method provided in this embodiment of the invention specifically includes the following steps: S1: Perform data augmentation on the cervical tissue OCT images collected by the client and group them into training set, validation set and test set; S2: Construct a Deep Evidence Learning (DEL) network architecture on the client side, including an image encoder and an evidence classifier. The image encoder uses ResNet18 to extract features from cervical OCT images (output shape is [B, K, H', W'], where K is the number of channels and H' / W' is the feature map space size). The evidence classifier converts the features into logits through a multi-layer fully connected network (output shape is [B, C], where C is the number of classes). The logits are further converted into evidence vectors through the Softplus activation function (evidence = F.softplus(logits), where F.softplus(logits) means that the softplus activation function is called through the functional module of PyTorch to convert the logits of the evidence classifier into a non-negative evidence vector, providing a basis for subsequent Dirichlet distribution parameter calculation and uncertainty modeling). The classification uncertainty is modeled based on the del distribution. S3: Design a dual knowledge distillation training framework, introduce probability knowledge distillation (Probability KD) and evidence knowledge distillation (Evidence KD) for collaborative optimization, and achieve model compression and performance improvement through a teacher-student network architecture, where both the teacher network and the student network are based on a deep evidence learning architecture; S4: Design an adaptive dynamic weight adjustment strategy, using a cosine annealing mechanism to dynamically adjust the knowledge distillation weights based on the number of training steps, thereby achieving an adaptive balance between knowledge distillation loss and hard classification loss during training. S5: Introduce interpretability loss to enhance feature learning, including In-Instance Latent Feature Knowledge (ILFK) loss and Inter-Class Variation Fuzzy Perception (IVFP) loss, to improve the model's interpretability of the intrinsic structure of features and class differences; S6: Design a cosine scheduler to adaptively adjust the learning rate and weight decay, and automatically optimize the training process according to the training progress. S7: Use the trained deep evidence learning model to classify and predict OCT images in the test set, and output the classification results and uncertainty estimates.

[0060] In one implementation, S2 specifically includes: S2.1: The image encoder uses ResNet18 as the backbone network and extracts the spatial structural features of cervical OCT images through multiple convolutional layers. The ResNet18 encoder outputs a 512-dimensional feature vector; S2.2: The evidence classifier consists of a multi-layer fully connected network. Encoded features are input into the classifier, and the Softplus activation function is used to output an evidence vector. The evidence vector is converted into the parameter alpha of a del distribution by incrementing by 1, thus achieving uncertainty modeling.

[0061] S2.3: The classification probability is calculated based on the expectation of the Dirichlet distribution (normalized to (alpha)). The uncertainty assessment is based on the sum of (alpha) (reflecting the sufficiency of evidence), supporting the model's output of classification results while providing a prediction confidence assessment. High uncertainty in a prediction indicates that the model is not confident enough in its prediction of that sample, and can be used to identify difficult or outlier samples. This design allows the DEL architecture to output classification results and uncertainty estimates simultaneously, improving the model's reliability and interpretability.

[0062] In one implementation, the double knowledge distillation mechanism in S3 specifically includes: S3.1: Probability Knowledge Distillation (Probability KD): Temperature scaling and the Softmax function are applied to the logits of the teacher and student networks, respectively, to calculate the KL divergence loss. The temperature parameter T is used to soften the probability distribution, allowing the teacher network to provide richer soft label information. The formula is:

[0063] Where T is the temperature parameter (default 3.0), used for the softening probability distribution; s represents the logits output by the student network, t represents the logits output by the teacher network, and KL(· ·) represents the KL divergence, which measures the difference between two probability distributions; S3.2: Evidence Knowledge Distillation (EKD): This method aligns the evidence vectors of the teacher and student networks by applying temperature scaling and the Softmax function (KL model) or by directly calculating the mean squared error (MSE model). The KL model achieves alignment by softening the evidence distribution, while the MSE model directly aligns the numerical values ​​of the evidence vectors. The formula is:

[0064] in Temperature parameters for evidence distillation (default is the same as T); This represents the evidence vector output by the student network. This represents the evidence vector output by the teacher network; S3.3: Double Distillation Loss Combination: Combining probability KD and evidence KD using either an independent weighting mode or a complementary mode. The independent weighting mode allows setting the weights for probability KD separately ( ) and evidence KD weights ( ), complementary mode through Control the mixing ratio of the two losses. The formula is: (Independent weight mode) or (Complementary model) In one implementation, the adaptive dynamic weight adjustment strategy in S4 specifically includes: S4.1: A cosine annealing mechanism is used to dynamically adjust the knowledge distillation weights based on the number of training steps. The cosine annealing mechanism controls the changes in weights during training through a cosine function, allowing the student network to focus more on the soft objectives of the teacher model in the early stages of training, and more on the hard objectives of the ground truth labels in the later stages. The calculation formula is:

[0065]

[0066] in This represents the current number of training steps. Total training steps Indicates the weighting adjustment factor. This is the minimum weight for knowledge distillation (default 0.2). This represents the maximum weight for knowledge distillation (default 0.8). S4.2: Weight Change Curve: In the early stages of training (current_step = 0), cosine_factor = 1, w = (0.8), the model emphasizes the soft objective of the teacher model; during the middle of training (current_step = total_steps / 2), cosine_factor = 0, w = (0.2), the model focuses on hard targets with true labels; at the end of training (current_step = total_steps), cosine_factor = 1, w = (0.8), the model again emphasizes the soft objective of the teacher model, achieving a smooth weight transition; S4.3: Loss Function Combination in the Knowledge Distillation Stage: The knowledge distillation loss and hard classification loss are combined with dynamic weights. The hard classification loss uses cross-entropy loss to directly optimize the prediction of the true label. The formula is:

[0067] in is the cross-entropy loss, and w is the dynamically adjusted KD weight.

[0068] In one implementation, the interpretability loss in S5 specifically includes: S5.1: In-Instance Latent Feature Knowledge (ILFK) Loss: Derived from Channel Relationship Loss ( ) and spatial relationship loss ( The algorithm consists of several components used to learn the intrinsic structural patterns of the feature maps output by the encoder. Channel relation loss learns channel dependencies by calculating the cosine similarity matrix between the channels of the feature maps of the teacher and student networks. Spatial relation loss learns the spatial saliency distribution by calculating the average value of the feature maps along the channel dimension. The formula is:

[0069] in:

[0070]

[0071] Where K is the number of channels, and H and W are the feature map space dimensions; The channel relationship matrix of the b-th sample in the teacher network is represented by the product of the normalized feature map and its transpose, which yields a [K×K] channel relationship matrix that reflects the cosine similarity between channels. Let B represent the channel relationship matrix of the b-th sample in the student network, where B represents the number of samples. It is the K-channel feature map of the b-th sample in the teacher network. The average value along the channel dimension is used to obtain the spatial saliency map of a single channel. ), It is the K-channel feature map of the b-th sample in the student network. The average value along the channel dimension is used to obtain the corresponding spatial saliency map. ).

[0072] S5.2: Inter-class Variation Fuzzy Perception (IVFP) Loss: This loss models the distribution of class features using style subsets, defining style centers (style_centers) and style radii (style_radii). Positive samples are rejected (away from all style centers), while negative samples are attracted (approaching the nearest style center). Sample labels are known ground truth labels from before training. Label source: Sample labels (label) are the true classification labels of the training data, already labeled before model training (e.g., the five-category labels for cervical OCT images: 0=normal, 1=mild inflammation, 2=cyst, 3=high-grade lesion, 4=cervical cancer). Style subset parameters are set using the `set_style_subsets` method, supporting dynamic updates based on the feature distribution of the teacher network, updated periodically. The formula is: For positive samples (label = 0):

[0073] Where C is the number of categories in the classification task, used to iterate through all categories to calculate the loss of positive samples. This design ensures that the IVFP loss can optimize the inter-class boundaries based on the true labels, thereby improving the classification performance and interpretability of the model. For negative samples (labels are 1-4):

[0074] in, This represents the distance between the i-th sample feature and the j-th style center. Let be the j-th style radius, and α and β be hyperparameters (default). =1.5, =0.2), The distance between the sample and the nearest style center. The radius of the nearest style center;

[0075] S5.3: Total Loss Function: Combines the interpretability loss with the base loss, controlling the contribution of the interpretability loss through weighting coefficients. ILFK Loss Weights ( The default value is 100, and the IVFP loss weights are ( The default value is 1, which can be adjusted according to task requirements. The formula is:

[0076] in The basic loss includes knowledge distillation loss and hard classification loss.

[0077] In one implementation, the cosine scheduler adaptive adjustment strategy in S6 specifically includes: S6.1: Cosine Learning Rate Scheduling: A cosine annealing strategy is used to automatically adjust the learning rate based on training progress. Cosine scheduling achieves a smooth decay of the learning rate through a cosine function, avoiding training instability caused by sudden changes in the learning rate. The formula is:

[0078] Where t is the current training step count, and T is the total training steps. Based on the learning rate, This is the minimum learning rate (the default is 1 / 100 of the base learning rate). S6.2: Warm-up Mechanism: A linear growth strategy is used in the early stages of training, increasing linearly from start_warmup_value (default 0) to base_value (base learning rate). The number of warm-up steps is warmup_epochs × niter_per_epoch. The warm-up mechanism helps the model converge stably in the early stages of training and avoids training oscillations caused by an excessively large learning rate. S6.3: Weight Decay Scheduling: Employs the same cosine scheduling strategy as the learning rate, automatically adjusting the weight decay coefficient based on the training progress. Weights decay from the initial value (weight_decay) to the final value (weight_decay_end) according to a cosine function, which helps in fine-tuning the model in the later stages of training.

[0079] In one implementation, a dynamic data augmentation strategy is employed during the training phase, including random horizontal flipping, rotation, color jittering, and grayscale conversion, to improve the model's generalization ability, and the images are uniformly scaled to a specific resolution.

[0080] In one implementation, the training process employs a cross-validation strategy, dividing the dataset into multiple folds, with each fold trained and evaluated independently. The final result is the average of the multiple folds, thereby reducing the variance of the model evaluation.

[0081] In one implementation, the programming language used is Python, the software tool used is PyTorch, and GPU-accelerated training is supported.

[0082] In one implementation, the classification result is a multi-class label (e.g., five categories: mild inflammation, cyst, eversion, high-grade lesion, cervical cancer). The model performs multi-class prediction for each test image, and the prediction result is output after being converted from evidence vectors into classification probabilities. At the same time, uncertainty estimates are provided to assess the prediction confidence.

[0083] In one implementation, the model supports evaluation on internal and external test sets (Huaxi dataset, Xiangya dataset, etc.), calculating accuracy (ACC), five-class F1 score, sensitivity, and specificity to evaluate the model's generalization ability and anomaly detection performance. Experimental results are shown in Table 1. Table 1 Experimental Results

[0084] In summary, the method provided by the embodiments of the present invention has the following beneficial effects: 1. Dual knowledge distillation mechanism: By co-optimizing probabilistic knowledge distillation and evidential knowledge distillation, the advantages of deep evidence learning models are fully utilized to achieve model compression and performance improvement.

[0085] 2. Adaptive weight adjustment: The cosine annealing mechanism is used to dynamically adjust the knowledge distillation weights based on the number of training steps, so as to achieve adaptive balance of loss during training and achieve stable optimization of the training process without manual parameter tuning.

[0086] 3. Enhanced interpretability: By using In-Instance Latent Feature Knowledge (ILFK) loss and Inter-Class Variation Fuzzy Perception (IVFP) loss, the interpretability of the model for the intrinsic structure of features and class differences is improved, meeting the clinical needs for diagnostic evidence.

[0087] 4. Uncertainty Quantification: Based on a deep evidence learning framework, it supports providing prediction confidence assessment while outputting classification results, which helps to identify difficult or abnormal samples.

[0088] 5. Adaptive learning rate scheduling: A cosine scheduler is used to automatically adjust the learning rate and weight decay according to the training progress, combined with a warm-up mechanism to achieve stable model convergence and fine-tuning.

[0089] Secondly, embodiments of the present invention also provide a cervical OCT image intelligent diagnostic system, including a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute a cervical OCT image intelligent diagnostic method as described in the above technical solution.

[0090] Thirdly, embodiments of the present invention also provide a computer-readable storage medium, including a readable storage medium on which a computer program is stored, wherein when the computer program is executed, it implements a cervical OCT image intelligent diagnosis method as described in the above technical solution.

[0091] The above embodiments are described only to clearly illustrate the basic technical solution of the present invention, but the present invention is not limited to the above embodiments; those skilled in the art to which the present invention pertains can make various modifications or additions to the specific embodiments described or use similar methods to replace them, but will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims; any simple modifications, equivalent changes and alterations made to the embodiments based on the technical essence of the present invention fall within the protection scope of the technical solution of the present invention.

Claims

1. A smart diagnostic method for cervical OCT images, characterized in that, include: Acquire cervical OCT images and perform data augmentation; A deep evidence learning network architecture is constructed, including an image encoder and an evidence classifier. The image encoder is used to extract features from cervical OCT images, and the evidence classifier converts the features into raw output values ​​logits and evidence vectors, and models the classification uncertainty based on the Dirichlet distribution. A dual knowledge distillation training framework is designed, which introduces probabilistic knowledge distillation and evidence knowledge distillation for collaborative optimization, thereby achieving network architecture compression and performance improvement. Both the teacher network and the student network are based on a deep evidence learning architecture. An adaptive dynamic weight adjustment strategy is designed, which uses a cosine annealing mechanism to dynamically adjust the knowledge distillation weights according to the number of training steps, balancing the knowledge distillation loss and the hard classification loss, thereby obtaining the base loss. Interpretability loss is introduced to enhance feature learning, including intra-instance latent feature knowledge loss and inter-class variation fuzzy perception loss, and combined with the base loss to construct the final total loss; The cosine scheduler is designed to adaptively adjust the learning rate and weight decay, and automatically optimize the overall training process according to the training progress. The trained deep evidence learning model is used to classify and predict the OCT images to be tested, and the classification results and uncertainty estimates are output.

2. The intelligent diagnostic method for cervical OCT images as described in claim 1, characterized in that: The image encoder uses ResNet18 to extract features from cervical OCT images. The evidence classifier converts the features into logits through a multi-layer fully connected network. The logits are then used to output an evidence vector through the Softplus activation function. The evidence vector is converted into the parameters α of the Dirichlet distribution by adding 1, thus realizing uncertainty modeling.

3. The intelligent diagnostic method for cervical OCT images as described in claim 1, characterized in that: The formula for calculating probabilistic knowledge distillation is: Where T is the temperature parameter for probabilistic knowledge distillation, used to soften the probability distribution; s represents the logits output by the student network, t represents the logits output by the teacher network, and KL(· ·) represents the KL divergence, which measures the difference between two probability distributions; For activation functions; The formula for calculating evidence knowledge distillation is: in Temperature parameters for distillation of evidence knowledge; This represents the evidence vector output by the student network. KL(·) represents the evidence vector output by the teacher network. ·) represents the KL divergence, which measures the difference between two probability distributions; For activation functions; Knowledge distillation loss is obtained by combining probabilistic knowledge distillation and evidential knowledge distillation using independent weighting patterns or complementary patterns.

4. The intelligent diagnostic method for cervical OCT images as described in claim 1, characterized in that: The formula for calculating the knowledge distillation weight is: in This represents the current number of training steps. Total training steps Indicates the weighting adjustment factor. This represents the minimum weight for knowledge distillation. This represents the maximum weight for knowledge distillation. Balancing the knowledge distillation loss and hard classification loss is achieved through the following formula: in, Basic loss, hard classification loss The loss is cross-entropy, and w is the dynamically adjusted knowledge distillation weight. This is due to knowledge distillation loss.

5. The intelligent diagnostic method for cervical OCT images as described in claim 1, characterized in that: In-instance latent feature knowledge loss Loss due to channel relationship Loss of spatial relationship The system comprises three components: a feature map for learning the intrinsic structural patterns of the encoder output; a channel relation loss for learning channel dependencies by calculating the cosine similarity matrix between channels of the feature map; and a spatial relation loss for learning the spatial saliency distribution by calculating the average value of the feature map along the channel dimension. The specific calculation formula is as follows: in: Where K is the number of channels, and H and W are the feature map space dimensions; This represents the channel relationship matrix of the b-th sample in the teacher network. Let B represent the channel relationship matrix of the b-th sample in the student network, where B represents the number of samples. The spatial saliency map of a single channel is obtained by averaging the K channel feature maps of the b-th sample in the teacher network along the channel dimension. The spatial saliency map is obtained by averaging the K channel feature maps of the b-th sample in the student network along the channel dimension.

6. The intelligent diagnostic method for cervical OCT images as described in claim 1, characterized in that: Fuzzy perception of inter-class changes The loss is modeled using style subsets to represent the categorical feature distribution. Style centers and styles are defined, and style subset parameters are set using the `set_style_subsets` method, supporting dynamic updates. These are periodically updated based on the feature distribution of the teacher network, as shown in the formula: For positive samples: Where C is the number of categories in the classification task; For negative samples: in This represents the distance between the i-th sample feature and the j-th style center. Let be the j-th style radius, and α and β be hyperparameters. The distance between the sample and the nearest style center. The radius of the nearest style center; 。 7. The intelligent diagnostic method for cervical OCT images as described in claim 1, characterized in that: The final total loss is: in Based on the loss, and These are the in-instance latent feature knowledge loss and the inter-class change fuzzy perception loss, respectively. and These are the weighting coefficients.

8. The intelligent diagnostic method for cervical OCT images as described in claim 1, characterized in that: The learning rate is set as follows: Where t is the current training step count, and T is the total training steps. Based on the learning rate, Minimum learning rate; A warm-up mechanism is used during training, and the same cosine scheduling strategy as the learning rate is adopted. The weight decay coefficient is automatically adjusted according to the training progress, and the weight decays from the initial value to the final value according to the cosine function.

9. A cervical OCT image intelligent diagnostic system, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the intelligent diagnostic method for cervical OCT images as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The method includes a readable storage medium on which a computer program is stored, and when the computer program is executed, it implements the intelligent diagnostic method for cervical OCT images as described in any one of claims 1-8.

Images