Self-adaptive diagnosis model training method and system for OCT (Optical Coherence Tomography) image

Abstract

The invention discloses a self-adaptive diagnosis model training method and system for OCT images. The system comprises an image data preprocessing module, a feature self-adaptive extraction module, a diagnosis model self-adaptive training module, a model performance dynamic evaluation module, a diagnosis reasoning and feedback module and a model iterative optimization module. According to the method, accurate adaptation of OCT images with different qualities and different equipment sources is realized through a hierarchical preprocessing and dynamic feature fusion mechanism, and heterogeneity interference of the images is effectively eliminated; based on hierarchical training of lesion complexity and dynamic loss function design, the model can focus on samples difficult to recognize and minority classes of lesions, and the diagnosis balance is improved.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence diagnostic technology, specifically to an adaptive diagnostic model training method and system for OCT images. Background Technology

[0002] OCT imaging, due to its ability to clearly reveal the microscopic structures of tissues such as the eye, has become a crucial basis for clinical diagnosis. Current OCT diagnostic model training suffers from the following problems: the use of fixed image preprocessing and feature extraction strategies fails to adapt to image heterogeneity caused by different devices and lesion types; the model training process lacks a dynamic adjustment mechanism, resulting in insufficient adaptability to differences in sample distribution; and training is disconnected from clinical application, making it difficult to optimize the model based on actual diagnostic feedback. These problems lead to weak model generalization ability, making it difficult to meet the diagnostic needs of complex clinical scenarios. Therefore, there is an urgent need for an adaptive diagnostic model training method and system for OCT imaging. Summary of the Invention

[0003] The purpose of this invention is to provide an adaptive diagnostic model training method and system for OCT images, so as to solve the problems existing in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an adaptive diagnostic model training system for OCT images, comprising an image data preprocessing module, a feature adaptive extraction module, a diagnostic model adaptive training module, a model performance dynamic evaluation module, a diagnostic inference and feedback module, and a model iterative optimization module; The image data preprocessing module performs quality grading, noise suppression and standardization processing on the original OCT images, and outputs standardized images with uniform pixel size. The adaptive feature extraction module includes a manual feature extraction unit, a deep feature extraction unit, and a feature fusion unit. The manual feature extraction unit extracts the texture, morphology, and interlayer structure features of the image. The deep feature extraction unit extracts deep semantic features through a pre-trained convolutional neural network. The feature fusion unit uses a dynamic weighting strategy to fuse the two types of features. The diagnostic model adaptive training module includes an initial model building unit, a dynamic loss calculation unit, and a hierarchical training unit. The initial model building unit builds a basic diagnostic network based on a residual network. The dynamic loss calculation unit introduces a sample difficulty coefficient and a class balance factor to calculate the loss value. The hierarchical training unit executes the training process hierarchically according to the complexity of the lesion. The model performance dynamic evaluation module includes an index calculation unit, an evaluation result analysis unit, and a threshold determination unit. The index calculation unit calculates diagnostic accuracy, sensitivity, and generalization indicators. The evaluation result analysis unit compares the differences between the indicators and preset thresholds. The threshold determination unit outputs model adjustment instructions. The diagnostic reasoning and feedback module includes an image input unit, a reasoning execution unit, and a feedback acquisition unit. The image input unit receives the OCT image to be diagnosed, the reasoning execution unit calls the trained model to output the diagnostic result, and the feedback acquisition unit collects the clinical diagnosis result and the doctor's correction opinion. The model iteration optimization module includes a feedback data processing unit, an incremental training unit, and a model update unit. The feedback data processing unit labels and classifies the collected feedback data, the incremental training unit performs incremental training of the model based on the new data, and the model update unit replaces the original model parameters and saves the new version.

[0005] Preferably, the image data preprocessing module specifically includes a quality assessment unit, a grading processing unit, and a standardization unit; Quality assessment unit: By calculating the signal-to-noise ratio, edge sharpness, and layer structure integrity of the image, OCT images are divided into three levels: high quality, medium quality, and low quality. The hierarchical processing unit uses a multi-scale nonlocal mean algorithm to suppress noise and artifacts in low-quality images, adaptive histogram equalization to enhance the contrast of lesion areas in medium-quality images, and Gaussian filtering to preserve details and smooth textures in high-quality images. Standardization Unit: The processed image is uniformly scaled to a preset pixel size, and the pixel grayscale value is standardized using the Z-score normalization method to eliminate the influence of grayscale differences between devices.

[0006] Preferably, the feature fusion unit of the adaptive feature extraction module adopts a dynamic weight fusion formula:

[0007] in, This is the fused adaptive feature vector; These are the manual feature weighting coefficients. These are the weight coefficients for deep features, and ; The feature vector is a handmade feature vector, which includes texture features, morphological features, and interlayer structure features. Texture features are gray-level co-occurrence matrix statistics, morphological features are the area and perimeter of the lesion area, and interlayer structure features are the thickness difference between the layers of the retina. This is a deep feature vector, output by the last fully connected layer of the convolutional neural network; and The feature contribution is obtained through feature contribution calculation, which is determined by the random forest algorithm to score the diagnostic importance of the two types of features.

[0008] Preferably, the dynamic loss calculation unit of the diagnostic model adaptive training module adopts the dynamic weighted cross-entropy loss formula:

[0009] in, This is the dynamic loss value; The number of samples in a single training session; This is the sample index, with a value range of 1 to... ; For the first The difficulty coefficient of each sample is calculated from the feature complexity and annotation reliability of the sample. The higher the feature complexity and the lower the annotation reliability, the better. The larger the value; For the first The category balance factor for lesions is obtained by normalizing the inverse of the number of samples in each category in the training set; For the first The sample corresponds to the first The true label for the type of lesion, with a value of 1 (belonging to this type) or 0 (not belonging to this type); For the model to predict the first The sample belongs to the first The probability of developing a lesion.

[0010] Preferably, the threshold determination unit of the model performance dynamic evaluation module presets three-level threshold standards: when the core indicators (accuracy and sensitivity) are both higher than the first-level threshold, the model continues the current training; when the core indicators are lower than the first-level threshold but higher than the second-level threshold, an instruction to adjust the learning rate is output; when the core indicators are lower than the second-level threshold or the generalization indicator is lower than the third-level threshold, an instruction to restart feature extraction and model training is output.

[0011] Preferably, the incremental training unit of the model iterative optimization module adopts a freeze-fine-tuning strategy: freezing the parameters of the first half of the convolutional layers of the basic diagnostic network and fine-tuning only the parameters of the second half of the convolutional layers and fully connected layers; the batch size and training epochs of incremental training are adaptively adjusted according to the number of feedback data samples, and the more feedback samples there are, the more training epochs are added accordingly.

[0012] Preferably, the image input unit is responsible for receiving the OCT image data to be diagnosed. During the receiving process, it performs preliminary verification on the basic information such as the image format and resolution to ensure that the input data meets the model processing requirements. Inference Execution Unit: After the image input unit completes data processing, the inference execution unit is triggered. It calls the pre-trained adaptive diagnostic model for OCT images, performs in-depth analysis and feature extraction on the input image data based on the model's internal algorithms and parameters, and outputs preliminary diagnostic results through calculation and logical judgment. Feedback Acquisition Unit: After the inference execution unit outputs the diagnostic results, the feedback acquisition unit starts working; it integrates the acquired information to form a feedback dataset.

[0013] A training method for an adaptive diagnostic model training system for OCT images includes the following steps: Step 1: System Initialization and Parameter Configuration Start each module of the system, preset the evaluation thresholds at each level in the model performance dynamic evaluation module, configure the basic network structure parameters and initial learning rate in the diagnostic model adaptive training module, load the pre-trained convolutional neural network weights in the feature adaptive extraction module, and enter the standby state after the system completes initialization. Step 2: OCT image data preprocessing The image data preprocessing module receives raw OCT images from multiple devices; the quality assessment unit calculates image quality indicators and classifies quality levels; the graded processing unit performs targeted denoising and enhancement processing according to the level; and the standardization unit scales the processed images to a uniform size and performs grayscale normalization, outputting a standardized image dataset. Step 3: Adaptive Feature Extraction The adaptive feature extraction module receives standardized images, and the manual feature extraction unit extracts texture, morphology and inter-layer structure features and constructs manual feature vectors. The deep feature extraction unit inputs the images into a pre-trained convolutional neural network to extract deep semantic features and form deep feature vectors. The feature fusion unit calculates feature contribution based on the random forest algorithm and generates an adaptive feature set through a dynamic weight fusion formula. Step 4: Adaptive Training of the Diagnostic Model The initial model building unit of the adaptive training module for the diagnostic model is based on a residual network to build a basic diagnostic model. The adaptive feature set is divided into three levels according to the complexity of the lesion: simple, medium, and complex. The hierarchical training unit first uses simple level samples to train the basic parameters of the model, and then gradually adds medium and complex level samples. The dynamic loss calculation unit calculates the loss value in each round of training using the dynamic weighted cross-entropy loss formula and uses the backpropagation algorithm to update the model parameters until the preset training rounds are reached. Step 5: Dynamic evaluation and adjustment of model performance The dynamic evaluation module for model performance tests the trained model by calculating accuracy, sensitivity, and generalization metrics. The evaluation result analysis unit compares the metrics with preset thresholds. If the metrics are higher than the first-level threshold, the model enters a state of pending diagnosis. If the metrics are lower than the first-level threshold but higher than the second-level threshold, the learning rate is adjusted and the system returns to step 4 to continue training. If the metrics are lower than the second-level threshold or the generalization is not up to standard, the system returns to step 3 to re-execute feature extraction and training. Step 6: Diagnostic Reasoning and Feedback Collection The image input unit of the diagnostic reasoning and feedback module receives clinical OCT images to be diagnosed. After preprocessing in step 2 and feature extraction in step 3, the reasoning execution unit calls the trained model to output the lesion type and diagnostic basis. The feedback acquisition unit collects the confirmed diagnosis results and correction opinions of doctors through the clinical interface, and associates the corresponding images and feature data to form a feedback dataset. Step 7: Iterative Model Optimization The feedback data processing unit of the model iteration optimization module labels and classifies the feedback dataset, and selects effective training samples; the incremental training unit adopts the freeze-fine-tuning strategy to perform incremental training on the original model based on the effective samples; the model update unit saves the optimized model parameters, replaces the original model and records the iteration log, and returns to step 4 after completing one iteration to wait for the next round of feedback data.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves precise adaptation to OCT images of different qualities and from different equipment sources through a hierarchical preprocessing and dynamic feature fusion mechanism, effectively eliminating image heterogeneity interference. Layered training based on lesion complexity and dynamic loss function design enable the model to focus on difficult-to-identify samples and minority lesions, improving diagnostic balance. A closed-loop mechanism for training, evaluation, inference, feedback, and iteration is constructed, allowing the model to continuously absorb clinical experience and adapt to ever-changing diagnostic needs. The adaptive fusion of handcrafted features and deep features preserves the anatomical features of OCT images while uncovering deep semantic relationships, maximizing feature value. Attached Figure Description

[0015] Figure 1 This is a system module diagram of the present invention; Figure 2 This is a schematic diagram of the feature adaptive extraction module of the present invention; Figure 3 This is a schematic diagram of the model performance dynamic evaluation module of the present invention; Figure 4 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1-4This invention provides an adaptive diagnostic model training system for OCT images, including an image data preprocessing module, an adaptive feature extraction module, an adaptive diagnostic model training module, a dynamic model performance evaluation module, a diagnostic inference and feedback module, and a model iterative optimization module. The image data preprocessing module performs quality grading, noise suppression, and standardization on the raw OCT images, and outputs standardized images with uniform pixel size. The adaptive feature extraction module includes a manual feature extraction unit, a deep feature extraction unit, and a feature fusion unit. The manual feature extraction unit extracts the texture, morphology, and interlayer structure features of the image. The deep feature extraction unit extracts deep semantic features through a pre-trained convolutional neural network. The feature fusion unit uses a dynamic weight strategy to fuse the two types of features. The diagnostic model adaptive training module includes an initial model building unit, a dynamic loss calculation unit, and a hierarchical training unit. The initial model building unit builds a basic diagnostic network based on the residual network. The dynamic loss calculation unit introduces the sample difficulty coefficient and the class balance factor to calculate the loss value. The hierarchical training unit executes the training process hierarchically according to the complexity of the lesion. The model performance dynamic evaluation module includes an index calculation unit, an evaluation result analysis unit, and a threshold determination unit. The index calculation unit calculates the diagnostic accuracy, sensitivity, and generalization indices. The evaluation result analysis unit compares the differences between the indices and preset thresholds. The threshold determination unit outputs model adjustment instructions. The diagnostic reasoning and feedback module includes an image input unit, a reasoning execution unit, and a feedback acquisition unit. The image input unit receives the OCT image to be diagnosed, the reasoning execution unit calls the trained model to output the diagnostic result, and the feedback acquisition unit collects the clinical diagnosis result and the doctor's correction opinion. The model iteration and optimization module includes a feedback data processing unit, an incremental training unit, and a model update unit. The feedback data processing unit labels and classifies the collected feedback data, the incremental training unit performs incremental training of the model based on the new data, and the model update unit replaces the original model parameters and saves the new version.

[0018] The image data preprocessing module specifically includes a quality assessment unit, a grading processing unit, and a standardization unit; Quality assessment unit: By calculating the signal-to-noise ratio, edge sharpness, and layer structure integrity of the image, OCT images are divided into three levels: high quality, medium quality, and low quality. The hierarchical processing unit uses a multi-scale nonlocal mean algorithm to suppress noise and artifacts in low-quality images, adaptive histogram equalization to enhance the contrast of lesion areas in medium-quality images, and Gaussian filtering to preserve details and smooth textures in high-quality images. Standardization Unit: The processed image is uniformly scaled to a preset pixel size, and the pixel grayscale value is standardized using the Z-score normalization method to eliminate the influence of grayscale differences between devices.

[0019] The feature fusion unit of the adaptive feature extraction module adopts a dynamic weight fusion formula:

[0020] in, This is the fused adaptive feature vector; These are the manual feature weighting coefficients. These are the weight coefficients for deep features, and ; The feature vector is a handmade feature vector, which includes texture features, morphological features, and interlayer structure features. Texture features are gray-level co-occurrence matrix statistics, morphological features are the area and perimeter of the lesion area, and interlayer structure features are the thickness difference between the layers of the retina. This is a deep feature vector, output by the last fully connected layer of the convolutional neural network; and The feature contribution is obtained through feature contribution calculation, which is determined by the random forest algorithm to score the diagnostic importance of the two types of features.

[0021] The dynamic loss calculation unit of the adaptive training module of the diagnostic model uses the dynamic weighted cross-entropy loss formula:

[0022] in, This is the dynamic loss value; The number of samples in a single training session; This is the sample index, with a value range of 1 to... ; For the first The difficulty coefficient of each sample is calculated from the feature complexity and annotation reliability of the sample. The higher the feature complexity and the lower the annotation reliability, the better. The larger the value; For the first The category balance factor for lesions is obtained by normalizing the inverse of the number of samples in each category in the training set; For the first The sample corresponds to the first The true label for the type of lesion, with a value of 1 (belonging to this type) or 0 (not belonging to this type); For the model to predict the first The sample belongs to the first The probability of developing a lesion.

[0023] The threshold determination unit of the model performance dynamic evaluation module presets three levels of threshold standards: when the core indicators (accuracy and sensitivity) are both higher than the first-level threshold, the model continues the current training; when the core indicators are lower than the first-level threshold but higher than the second-level threshold, an instruction to adjust the learning rate is output; when the core indicators are lower than the second-level threshold or the generalization indicator is lower than the third-level threshold, an instruction to restart feature extraction and model training is output.

[0024] The incremental training unit of the model iteration optimization module adopts a freeze-fine-tuning strategy: the parameters of the first half of the convolutional layers of the basic diagnostic network are frozen, and only the parameters of the second half of the convolutional layers and fully connected layers are fine-tuned; the batch size and training epochs of incremental training are adaptively adjusted according to the number of feedback data samples. The more feedback samples there are, the more training epochs are added accordingly.

[0025] Image input unit: Responsible for receiving OCT image data to be diagnosed. During the receiving process, it will perform preliminary verification of basic information such as image format and resolution to ensure that the input data meets the model processing requirements. Inference Execution Unit: After the image input unit completes data processing, the inference execution unit is triggered. It calls the pre-trained adaptive diagnostic model for OCT images, performs in-depth analysis and feature extraction on the input image data based on the model's internal algorithms and parameters, and outputs preliminary diagnostic results through calculation and logical judgment. Feedback Acquisition Unit: After the inference execution unit outputs the diagnostic results, the feedback acquisition unit starts working; it integrates the acquired information to form a feedback dataset.

[0026] A training method for an adaptive diagnostic model training system for OCT images includes the following steps: Step 1: System Initialization and Parameter Configuration Start each module of the system, preset the evaluation thresholds at each level in the model performance dynamic evaluation module, configure the basic network structure parameters and initial learning rate in the diagnostic model adaptive training module, load the pre-trained convolutional neural network weights in the feature adaptive extraction module, and enter the standby state after the system completes initialization. Step 2: OCT image data preprocessing The image data preprocessing module receives raw OCT images from multiple devices; the quality assessment unit calculates image quality indicators and classifies quality levels; the graded processing unit performs targeted denoising and enhancement processing according to the level; and the standardization unit scales the processed images to a uniform size and performs grayscale normalization, outputting a standardized image dataset. Step 3: Adaptive Feature Extraction The adaptive feature extraction module receives standardized images, and the manual feature extraction unit extracts texture, morphology and inter-layer structure features and constructs manual feature vectors. The deep feature extraction unit inputs the images into a pre-trained convolutional neural network to extract deep semantic features and form deep feature vectors. The feature fusion unit calculates feature contribution based on the random forest algorithm and generates an adaptive feature set through a dynamic weight fusion formula. Step 4: Adaptive Training of the Diagnostic Model The initial model building unit of the adaptive training module for the diagnostic model is based on a residual network to build a basic diagnostic model. The adaptive feature set is divided into three levels according to the complexity of the lesion: simple, medium, and complex. The hierarchical training unit first uses simple level samples to train the basic parameters of the model, and then gradually adds medium and complex level samples. The dynamic loss calculation unit calculates the loss value in each round of training using the dynamic weighted cross-entropy loss formula and uses the backpropagation algorithm to update the model parameters until the preset training rounds are reached. Step 5: Dynamic evaluation and adjustment of model performance The dynamic evaluation module for model performance tests the trained model by calculating accuracy, sensitivity, and generalization metrics. The evaluation result analysis unit compares the metrics with preset thresholds. If the metrics are higher than the first-level threshold, the model enters a state of pending diagnosis. If the metrics are lower than the first-level threshold but higher than the second-level threshold, the learning rate is adjusted and the system returns to step 4 to continue training. If the metrics are lower than the second-level threshold or the generalization is not up to standard, the system returns to step 3 to re-execute feature extraction and training. Step 6: Diagnostic Reasoning and Feedback Collection The image input unit of the diagnostic reasoning and feedback module receives clinical OCT images to be diagnosed. After preprocessing in step 2 and feature extraction in step 3, the reasoning execution unit calls the trained model to output the lesion type and diagnostic basis. The feedback acquisition unit collects the confirmed diagnosis results and correction opinions of doctors through the clinical interface, and associates the corresponding images and feature data to form a feedback dataset. Step 7: Iterative Model Optimization The feedback data processing unit of the model iteration optimization module labels and classifies the feedback dataset, and selects effective training samples; the incremental training unit adopts the freeze-fine-tuning strategy to perform incremental training on the original model based on the effective samples; the model update unit saves the optimized model parameters, replaces the original model and records the iteration log, and returns to step 4 after completing one iteration to wait for the next round of feedback data.

[0027] Example: Taking the OCT image diagnosis scenario of an ophthalmology department in a top-tier hospital as an example, the system first completes the initial configuration after deployment: The model performance dynamic evaluation module presets an accuracy of ≥92% and a sensitivity of ≥90% as the first-level threshold, and 85%-89% as the second-level threshold; The diagnostic model adaptive training module uses ResNet50 as the base network with an initial learning rate of 0.001; the feature adaptive extraction module loads the weights of the VGG16 network pre-trained on the ImageNet dataset.

[0028] On the first day of system operation, a total of 127 raw image data were received from three different brands of OCT equipment. The image data preprocessing module found the following quality assessment results: 68 high-quality images (signal-to-noise ratio > 25dB, edge sharpness > 0.8, layer structure integrity score ≥ 4 points), 42 medium-quality images (index range 20-25dB, 0.6-0.8, 3-4 points), and 17 low-quality images (< 20dB, < 0.6, < 3 points).

[0029] The hierarchical processing unit employs a 7×7 window multi-scale nonlocal mean algorithm for low-quality images, adaptive histogram equalization of 128×128 regions for medium-quality images, and 3×3 Gaussian filtering for high-quality images. =1.2), all images are uniformly scaled to 512×512 pixels using standardized units, and grayscale values ​​are standardized to a range of mean 0 and standard deviation 1 using Z-score.

[0030] In the adaptive feature extraction stage, the manual feature extraction unit extracted manual feature vectors from 68 high-quality images, including 16-dimensional gray-level co-occurrence matrix statistics, 3-dimensional morphological features (lesion area, perimeter, and roundness), and 5-dimensional interlayer structural features (retinal nerve fiber layer, ganglion cell layer, inner plexiform layer, outer plexiform layer, and retinal pigment epithelium thickness). The deep feature extraction unit outputs a 4096-dimensional deep feature vector through the fully connected layers of a pre-trained VGG16 network; the feature fusion unit calculates the handcrafted feature weight coefficients based on the random forest algorithm. =0.38, depth feature weight coefficient 0.62, a 512-dimensional adaptive feature set is generated through a dynamic weight fusion formula.

[0031] During adaptive training of the diagnostic model, the initial model building unit constructs a residual network containing 50 convolutional layers. The hierarchical training unit divides the 127 adaptive feature sets into simple level (42 cases of simple macular edema), medium level (58 cases of macular lesions with vitreous traction), and complex level (27 cases of diabetic retinopathy with epiretinal membrane) according to the complexity of the lesions. The parameters of the first 30 layers of the model are trained first using simple level samples (20 training rounds), and then medium and complex level samples are gradually added (15 more training rounds each). The dynamic loss calculation unit calculates the loss based on the sample difficulty coefficient in each training round. (Value range 0.7-1.3) and category balance factor (0.85 for macular edema, 0.92 for vitreous traction, and 1.15 for diabetic retinopathy) The dynamically weighted cross-entropy loss was calculated, and the Adam optimizer was used to update the model parameters. Dynamic evaluation of model performance showed that after 50 rounds of training, the model achieved an accuracy of 93.2%, a sensitivity of 91.5%, and a generalization index (cross-device diagnostic consistency) of 90.8% on the test set (26 cases), all of which were higher than the first-level threshold, and the system entered the waiting-for-diagnosis state.

[0032] During the first week of the diagnostic reasoning phase, the image input unit received 213 clinical OCT images for diagnosis. After preprocessing and feature extraction, the reasoning execution unit called the trained model to output diagnostic results. Among them, 198 cases were consistent with the doctor's diagnosis (including 142 cases of macular degeneration, 36 cases of retinal detachment, and 20 cases of glaucoma), and 15 cases required correction (10 cases were misdiagnosed as macular edema but were actually central serous chorioretinopathy, and 5 cases were missed in the diagnosis of diabetic retinopathy microaneurysms). The feedback acquisition unit combines these 15 related images and feature data to form a feedback dataset, which includes image quality indicators (12 medium quality and 3 low quality), feature vectors (mean deviation of manual features is 12.6%, and mean deviation of depth features is 8.3%), and doctor's correction opinions (marking the differences in inter-layer structural features of misdiagnosed areas and the deep semantic feature patterns of missed lesions).

[0033] During model iteration and optimization, the feedback data processing unit labeled and verified the feedback dataset, removed 3 invalid samples caused by image artifacts, and the remaining 12 cases were classified into central serous chorioretinopathy (7 cases) and diabetic retinopathy (5 cases) according to the lesion type. The incremental training unit adopts a strategy of freezing the parameters of the first 20 convolutional layers and fine-tuning the parameters of the last 30 convolutional layers and fully connected layers. Incremental training is performed based on 12 valid samples, with a batch size of 4 and 30 training rounds. The model update unit saves the optimized model parameters (weight update ratio 18.7%), replaces the original model, and records the iteration log, including the training sample distribution, parameter adjustment range, and performance improvement indicators. After completing the first iteration, it returns to step 4 to wait for the next round of feedback data.

[0034] The following week's operation showed that after training with 56 additional feedback data, the model's diagnostic accuracy increased to 94.7%, sensitivity increased to 93.1%, diagnostic specificity for central serous chorioretinopathy increased from 82% to 89%, and detection sensitivity for diabetic retinopathy microaneurysms increased from 76% to 85%. The system's ability to continuously adapt to clinical diagnostic needs was validated.

[0035] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive diagnostic model training system for OCT images, characterized in that: It includes an image data preprocessing module, an adaptive feature extraction module, an adaptive training module for the diagnostic model, a dynamic evaluation module for model performance, a diagnostic inference and feedback module, and a model iterative optimization module; The image data preprocessing module performs quality grading, noise suppression and standardization processing on the original OCT images, and outputs standardized images with uniform pixel size. The adaptive feature extraction module includes a manual feature extraction unit, a deep feature extraction unit, and a feature fusion unit. The manual feature extraction unit extracts the texture, morphology, and interlayer structure features of the image. The deep feature extraction unit extracts deep semantic features through a pre-trained convolutional neural network. The feature fusion unit uses a dynamic weighting strategy to fuse the two types of features. The diagnostic model adaptive training module includes an initial model building unit, a dynamic loss calculation unit, and a hierarchical training unit. The initial model building unit builds a basic diagnostic network based on a residual network. The dynamic loss calculation unit introduces a sample difficulty coefficient and a class balance factor to calculate the loss value. The hierarchical training unit executes the training process hierarchically according to the complexity of the lesion. The model performance dynamic evaluation module includes an index calculation unit, an evaluation result analysis unit, and a threshold determination unit. The index calculation unit calculates diagnostic accuracy, sensitivity, and generalization indicators. The evaluation result analysis unit compares the differences between the indicators and preset thresholds. The threshold determination unit outputs model adjustment instructions. The diagnostic reasoning and feedback module includes an image input unit, a reasoning execution unit, and a feedback acquisition unit. The image input unit receives the OCT image to be diagnosed, the reasoning execution unit calls the trained model to output the diagnostic result, and the feedback acquisition unit collects the clinical diagnosis result and the doctor's correction opinion. The model iteration optimization module includes a feedback data processing unit, an incremental training unit, and a model update unit. The feedback data processing unit labels and classifies the collected feedback data, the incremental training unit performs incremental training of the model based on the new data, and the model update unit replaces the original model parameters and saves the new version.

2. The adaptive diagnostic model training system for OCT images according to claim 1, characterized in that: The image data preprocessing module specifically includes a quality assessment unit, a grading processing unit, and a standardization unit; Quality assessment unit: By calculating the signal-to-noise ratio, edge sharpness, and layer structure integrity of the image, OCT images are divided into three levels: high quality, medium quality, and low quality. The hierarchical processing unit uses a multi-scale nonlocal mean algorithm to suppress noise and artifacts in low-quality images, adaptive histogram equalization to enhance the contrast of lesion areas in medium-quality images, and Gaussian filtering to preserve details and smooth textures in high-quality images. Standardization Unit: The processed image is uniformly scaled to a preset pixel size, and the pixel grayscale value is standardized using the Z-score normalization method to eliminate the influence of grayscale differences between devices.

3. The adaptive diagnostic model training system for OCT images according to claim 1, characterized in that: The feature fusion unit of the adaptive feature extraction module adopts a dynamic weight fusion formula: in, This is the fused adaptive feature vector; These are the manual feature weighting coefficients. These are the weight coefficients for deep features, and ; The feature vector is a handmade feature vector, which includes texture features, morphological features, and interlayer structure features. Texture features are gray-level co-occurrence matrix statistics, morphological features are the area and perimeter of the lesion area, and interlayer structure features are the thickness difference between the layers of the retina. This is a deep feature vector, output by the last fully connected layer of the convolutional neural network; and The feature contribution is obtained through feature contribution calculation, which is determined by the random forest algorithm to score the diagnostic importance of the two types of features.

4. The adaptive diagnostic model training system for OCT images according to claim 1, characterized in that: The dynamic loss calculation unit of the adaptive training module of the diagnostic model adopts the dynamic weighted cross-entropy loss formula: in, This is the dynamic loss value; The number of samples in a single training session; This is the sample index, with a value range of 1 to... ; For the first The difficulty coefficient of each sample is calculated from the feature complexity and annotation reliability of the sample. The higher the feature complexity and the lower the annotation reliability, the better. The larger the value; For the first The category balance factor for lesions is obtained by normalizing the inverse of the number of samples in each category in the training set; For the first The sample corresponds to the first The true label for the type of lesion, with a value of 1 or 0; For the model to predict the first The sample belongs to the first The probability of lesions.

5. The adaptive diagnostic model training system for OCT images according to claim 1, characterized in that: The threshold determination unit of the dynamic evaluation module for model performance has a preset three-level threshold standard: when all core indicators are higher than the first-level threshold, the model continues the current training; when the core indicators are lower than the first-level threshold but higher than the second-level threshold, an instruction to adjust the learning rate is output; when the core indicators are lower than the second-level threshold or the generalization indicator is lower than the third-level threshold, an instruction to restart feature extraction and model training is output.

6. The adaptive diagnostic model training system for OCT images according to claim 1, characterized in that: The incremental training unit of the model iterative optimization module adopts a freeze-fine-tuning strategy: the parameters of the first half of the convolutional layers of the basic diagnostic network are frozen, and only the parameters of the second half of the convolutional layers and fully connected layers are fine-tuned; the batch size and training epochs of incremental training are adaptively adjusted according to the number of feedback data samples, and the more feedback samples there are, the more training epochs are added accordingly.

7. The adaptive diagnostic model training system for OCT images according to claim 1, characterized in that: The image input unit is responsible for receiving the OCT image data to be diagnosed. During the receiving process, it performs preliminary verification on the basic information such as the image format and resolution to ensure that the input data meets the model processing requirements. Inference Execution Unit: After the image input unit completes data processing, the inference execution unit is triggered. It calls the pre-trained adaptive diagnostic model for OCT images, performs in-depth analysis and feature extraction on the input image data based on the model's internal algorithms and parameters, and outputs preliminary diagnostic results through calculation and logical judgment. Feedback Acquisition Unit: After the inference execution unit outputs the diagnostic results, the feedback acquisition unit starts working; it integrates the acquired information to form a feedback dataset.

8. The training method for an adaptive diagnostic model training system for OCT images according to any one of claims 1-7, characterized in that: Includes the following steps: Step 1: System Initialization and Parameter Configuration Start each module of the system, preset the evaluation thresholds at each level in the model performance dynamic evaluation module, configure the basic network structure parameters and initial learning rate in the diagnostic model adaptive training module, load the pre-trained convolutional neural network weights in the feature adaptive extraction module, and enter the standby state after the system completes initialization. Step 2: OCT image data preprocessing The image data preprocessing module receives raw OCT images from multiple devices, the quality assessment unit calculates image quality indicators and classifies quality levels, and the graded processing unit performs targeted denoising and enhancement processing according to the grade. The normalization unit scales the processed image to a uniform size and normalizes its grayscale, outputting a normalized image dataset. Step 3: Adaptive Feature Extraction The adaptive feature extraction module receives standardized images, while the manual feature extraction unit extracts texture, morphology, and interlayer structure features and constructs manual feature vectors. The deep feature extraction unit inputs the images into a pre-trained convolutional neural network to extract deep semantic features and form deep feature vectors. The fusion unit calculates feature contribution based on the random forest algorithm and generates an adaptive feature set through a dynamic weight fusion formula. Step 4: Adaptive Training of the Diagnostic Model The initial model building unit of the adaptive training module for the diagnostic model is based on a residual network to build a basic diagnostic model. The adaptive feature set is divided into three levels according to the complexity of the lesion: simple, medium, and complex. The hierarchical training unit first uses simple level samples to train the basic parameters of the model, and then gradually adds medium and complex level samples. The dynamic loss calculation unit calculates the loss value in each round of training using the dynamic weighted cross-entropy loss formula and uses the backpropagation algorithm to update the model parameters until the preset training rounds are reached. Step 5: Dynamic evaluation and adjustment of model performance The dynamic evaluation module for model performance tests the trained model by calculating accuracy, sensitivity, and generalization metrics. The evaluation result analysis unit compares the metrics with preset thresholds. If the metrics are higher than the first-level threshold, the model enters a state of pending diagnosis. If the metrics are lower than the first-level threshold but higher than the second-level threshold, the learning rate is adjusted and the system returns to step 4 to continue training. If the metrics are lower than the second-level threshold or the generalization is not up to standard, the system returns to step 3 to re-execute feature extraction and training. Step 6: Diagnostic Reasoning and Feedback Collection The image input unit of the diagnostic reasoning and feedback module receives clinical OCT images to be diagnosed. After preprocessing in step 2 and feature extraction in step 3, the reasoning execution unit calls the trained model to output the lesion type and diagnostic basis. The feedback acquisition unit collects the confirmed diagnosis results and correction opinions of doctors through the clinical interface, and associates the corresponding images and feature data to form a feedback dataset. Step 7: Iterative Model Optimization The feedback data processing unit of the model iteration and optimization module performs annotation verification and category classification on the feedback dataset and selects effective training samples. The incremental training unit employs a freeze-fine-tuning strategy, performing incremental training on the original model based on valid samples. The model update unit saves the optimized model parameters, replaces the original model, and records the iteration log. After completing one iteration, it returns to step 4 to wait for the next round of feedback data.

Images