Medical image intelligent diagnosis method and system based on deep learning

Abstract

The invention discloses a medical image intelligent diagnosis method and system based on deep learning, and relates to the technical field of medical iconography, and the method comprises the steps: receiving a multi-mechanism medical image with an initial diagnosis label, and generating an enhanced sample according with anatomical features through an adaptive data enhancement module based on an anatomical structure according to organ partitions; a hierarchical federated learning framework is established, a central server cooperates with multi-mechanism edge nodes, the edge nodes train an initial deep learning model by using local data, a central aggregation gradient generates global parameter feedback, and iteration is performed until the model is converged; constructing an annotation consistency evaluation module, carrying out secondary annotation by the specialist physician, calculating a weighted Kappa coefficient and a Dice similarity coefficient, and if the index does not reach the standard, rechecking by a secondary and primary physician; establishing an annotation-model feedback closed loop, and finely adjusting the local model by using recheck data; inputting a to-be-diagnosed image into the global model, outputting a lesion thermodynamic diagram, property probability distribution and confidence, and triggering manual review and outputting a feature analysis report when the confidence is low.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, specifically to a deep learning-based intelligent diagnostic method and system for medical images. Background Technology

[0002] Traditional medical imaging diagnosis relies on doctors manually interpreting images, which is inefficient when faced with massive amounts of image data, and the diagnostic results are significantly affected by the doctor's experience and fatigue level. Invasive biopsies have sampling biases and cause patient discomfort, while traditional image analysis struggles to quantify complex pathological features.

[0003] With breakthroughs in deep learning technology, convolutional neural networks (CNNs) can automatically extract features such as edges and textures from images. Generative AI (GAN / diffusion models) addresses the problem of sample scarcity by synthesizing image data. These models are deployed to terminals such as CT scanners and ultrasound equipment, enabling real-time diagnosis through edge computing and combining quantum algorithms to improve the efficiency of high-dimensional image processing.

[0004] The growing demand for precision medicine, including early cancer screening (such as mammography for breast cancer) and chronic disease management (such as diabetic retinopathy), urgently requires automated and high-precision diagnosis. AI-assisted diagnosis can alleviate the shortage of specialists in primary hospitals, achieve end-to-end intelligentization, and promote the downward flow of high-quality resources.

[0005] Therefore, there is an urgent need for a deep learning-based intelligent diagnostic method for medical images. This method faces challenges related to data quality and annotation, including data scarcity and bias. Medical images are subject to privacy protection restrictions, resulting in limited sample sizes in publicly available datasets and a severe lack of data on rare diseases. Furthermore, inconsistent annotations, such as differing judgments by different doctors regarding nodule boundaries and lesion characteristics, lead to inconsistent annotation standards and directly affect model training performance. Summary of the Invention

[0006] To address the aforementioned technical problems, a deep learning-based intelligent diagnostic method and system for medical images is provided. This technical solution solves the problems of data scarcity and bias, as well as inconsistent labeling.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based intelligent diagnostic method for medical images includes: It receives medical image data from multiple independent medical institutions, with each institution's image data accompanied by an initial diagnostic label; it performs privacy protection processing on the image data from each medical institution; and it constructs an adaptive data augmentation module based on anatomical structure to generate augmented samples that conform to the anatomical characteristics of the human organ corresponding to the image. A hierarchical federated learning framework is constructed, consisting of a central server node and multiple edge nodes in medical institutions. Each edge node deploys an initial deep learning model with the same structure. Each edge node trains the model using locally preprocessed image data. The central server node performs weighted aggregation of the gradients uploaded by each edge node, generates global model parameters after aggregation, and feeds them back to each edge node. The training is iterated until the model loss function converges. A multi-dimensional annotation consistency assessment module was constructed, in which multiple specialists with certain clinical experience performed secondary annotation on the same batch of image data. The annotation content included the boundary coordinates of the lesion area and the classification of the lesion nature. Annotation consistency index was calculated, including the weighted Kappa coefficient and the Dice similarity coefficient. If any index was lower than the preset threshold, the annotation review process was initiated, and the final judgment was made by a doctor with the title of associate chief physician or above. A closed-loop annotation-model feedback mechanism is established, using the reviewed and annotated data as a supplementary training set, which is then input into the edge nodes of the federated learning framework to fine-tune the local model and update the lesion recognition weights. Medical image data to be diagnosed is received and input into the trained global deep learning model. The model outputs a heatmap of the lesion region, the probability distribution of lesion characteristics, and diagnostic confidence. If the diagnostic confidence is lower than a predetermined proportion, a manual review reminder is automatically triggered, and a report on the model's analysis of lesion characteristics is output.

[0008] Preferably, the privacy protection processing of the image data from various medical institutions specifically includes: Differential privacy technology is used to inject noise of a preset strength into the identity field metadata containing the patient's name and medical record number; homomorphic encryption algorithm is used to encrypt the key feature data of the image's pixel features and texture features, and the encrypted data can be used in subsequent model training without decryption.

[0009] Preferably, the construction of the adaptive data augmentation module based on anatomical structure, which generates augmented samples that conform to the anatomical features of the human organ corresponding to the image, specifically includes: An anatomical database is established to structurally store information on the morphological characteristics, tissue density distribution patterns, and common lesion areas of different organs; the human organ partitions corresponding to the input images are automatically identified through image segmentation algorithms. Conventional enhanced sample generation scales the organ region to its original size based on organ boundary coordinates to simulate the image representation of changes in imaging distance or organ morphology; and adjusts the image grayscale values ​​by a predetermined ratio with reference to the grayscale range of normal and diseased tissues to simulate the grayscale differences of scanning parameters and pathological conditions. If a single institution has insufficient rare disease image samples, the generative adversarial network submodule is activated. Based on existing rare disease images, it learns the association pattern between lesions and background, and generates simulated samples with differences in core lesion features less than or equal to a predetermined proportion to supplement rare disease data.

[0010] Preferably, the construction of the hierarchical federated learning framework specifically includes: The central server node is configured with a parameter aggregation engine, a global model storage module, and node communication authentication. It is responsible for receiving model gradients uploaded by edge nodes, performing aggregation calculations, generating a global model, and distributing it to each edge node. At the same time, it records the model update logs of each node through blockchain technology. The edge nodes of the medical institution are each equipped with a local data storage unit, a model training engine and an encrypted communication module; each edge node is pre-deployed with an initial deep learning model with a consistent structure. The model input layer is adapted to the medical image modality, the hidden layer contains 3-5 residual blocks, and the output layer is a lesion classification probability vector.

[0011] Preferably, the model training at each edge node using locally preprocessed image data specifically includes: Edge nodes read preprocessed image data from local storage units and divide it into training and validation sets according to a predetermined ratio. The training set is grouped by organ type, and each group uses class-balanced sampling. Local training parameters are trained using the Adam optimizer, with batch size adjusted according to node computing power, and weighted cross-entropy is used as the loss function. After each round of local training is completed, the edge nodes calculate the gradient of the model parameters, encrypt the gradient using an asymmetric encryption algorithm, and transmit the encrypted gradient data only to the central server. The original images and intermediate training data are always stored locally.

[0012] Preferably, the central server node performs weighted aggregation of the gradients uploaded by each edge node, generates global model parameters after aggregation, and feeds them back to each edge node. Iterative training until the model loss function converges specifically includes: Gradient-weighted aggregation: After the central server decrypts the gradients of each edge node, it calculates the aggregation weights according to the formula; by calculating the cosine similarity between the gradient vector and the historical average gradient, if the similarity is less than a predetermined threshold, it is judged as an abnormal gradient and corrected by the truncated mean method. After aggregation, global model parameters are generated. After model compression, the parameters are encrypted and distributed to each edge node. The edge node updates its local model with the global parameters and repeats the local training, gradient uploading, and parameter update process until the loss function of the validation set decreases by less than a predetermined threshold for three consecutive rounds.

[0013] Preferably, the calculation of the annotation consistency index specifically includes: The weighted Kappa coefficient is calculated to assess the consistency of lesion classification. A contingency table is constructed based on the doctor's annotation results, and a clinical weight factor is introduced. The classification consistency is quantified by calculating the ratio of the actual consistency rate to the expected consistency rate. The Dice similarity coefficient calculation is based on the consistency of lesion boundaries. It extracts lesion areas labeled by different doctors, calculates the overlap by the ratio of twice the area intersection to the sum of the area pixels, and sets weights according to the difference in importance between the core area and the edge area of ​​the lesion to obtain the comprehensive Dice coefficient. Based on the accuracy requirements of medical diagnosis and industry consensus, preset qualification thresholds for two types of indicators were determined as the criteria for judging whether the labeling consistency meets the standards.

[0014] Preferably, the establishment of the annotation-model feedback closed loop specifically includes: Supplementary training set processing involves dividing the final labeled data after verification into a supplementary training subset and a validation subset according to a certain ratio, and associating them with corresponding image feature labels. The local model fine-tuning mechanism involves the federated learning edge nodes receiving supplementary training sets and then using mini-batch gradient descent to fine-tune the local model, reducing the learning rate to 1 / 5 of the initial training rate. The focus is on updating the weights of convolutional layers related to lesion identification in the model, and setting a 1.5-fold learning weight for the lesion features corrected in the verification annotation. After weight update verification, the model performance is evaluated by supplementing the verification subset after fine-tuning. If the lesion recognition accuracy improves by more than or equal to a predetermined proportion, the updated local model weights are retained and participate in the next round of federated aggregation; otherwise, the weights are backtracked to the weights before fine-tuning, and the difference features are recorded for optimization in the next supplementary training.

[0015] Preferably, the model outputs a heatmap of the lesion area, a probability distribution of the lesion nature, and a diagnostic confidence level, specifically including: Upon receiving the medical images to be diagnosed, the system automatically performs privacy-preserving and standardization processing on the same data source as the training data to ensure it matches the model's input requirements. The pre-processed images are then input into the trained global deep learning model. The model uses a feature extraction layer to locate potential lesion areas and a classification layer to output multi-dimensional results. The heatmap of the lesion area highlights high-probability lesion areas identified by the model with a color gradient, and the overlay of coordinate grid facilitates localization; the probability distribution of lesion nature outputs the predicted probabilities of benign, malignant, and suspected lesions in percentage form, with the highest probability item being the preliminary judgment result; it is calculated by combining the lesion feature matching degree and the model output probability entropy value, with the lower the entropy value, the higher the confidence level.

[0016] Furthermore, a deep learning-based intelligent medical imaging diagnostic system includes: Data acquisition and preprocessing module: used to acquire medical image data from multiple institutions, perform format standardization, noise removal and equipment deviation correction on the image data, and output preprocessed image data; A hierarchical federated learning framework consists of a central server node, which includes a parameter aggregation engine, a global model storage unit, and node communication authentication; and multiple medical institution edge nodes, each of which deploys a local model training engine, an encrypted communication unit, and a local data storage unit. The local model training engine is pre-configured with an initial deep learning model of the same structure. The annotation consistency assessment module includes: an annotation task allocation unit, which selects specialists with clinical experience to perform secondary annotation on the same batch of image data, including the boundary coordinates of lesion areas and the classification of lesion nature; a consistency calculation unit, which calculates the weighted Kappa coefficient and Dice similarity coefficient; and a review trigger unit, which initiates the annotation review process when the weighted Kappa coefficient and Dice similarity coefficient are lower than the predetermined threshold, calling the terminal devices of doctors with the title of associate chief physician or above to make the final judgment. The annotation-model feedback closed-loop module includes: a supplementary training set construction unit, which divides the verified annotation data into a supplementary training subset and a validation subset; and a local model fine-tuning unit, which controls the edge nodes to fine-tune the local model using mini-batch gradient descent, focusing on updating the weights of the convolutional layers related to lesion recognition. Intelligent diagnostic reasoning module: Image input unit, which receives medical image data to be diagnosed and performs privacy protection processing; Model reasoning unit, which inputs the preprocessed image into the trained global model and outputs a heat map of the lesion area, probability distribution of lesion nature, and diagnostic confidence level; Manual review triggering unit, which automatically pushes a review reminder to the doctor's terminal when the diagnostic confidence level is lower than a predetermined threshold, and outputs an auxiliary report containing lesion feature analysis and reference cases.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates multi-source data by receiving image data from multiple independent medical institutions with initial diagnostic labels, alleviating the problem of limited sample size in publicly available datasets. It employs differential privacy desensitization and homomorphic encryption techniques to ensure patient privacy while guaranteeing data usability for model training, thus overcoming limitations in medical image privacy protection. The constructed anatomical-based adaptive data augmentation module generates augmented samples conforming to anatomical features according to human organ partitions, avoiding feature distortion caused by general augmentation and effectively supplementing scarce data. Especially in the field of rare diseases where data is extremely scarce, precise augmentation can reduce data bias and provide a high-quality sample foundation for model training.

[0018] The central server collaborates with edge nodes of multiple medical institutions. Edge nodes train models based on local data and only upload gradients instead of raw data, which protects the privacy of institutional data and integrates the computing power and data resources of multiple institutions. The gradient weighted aggregation mechanism of the central server enables nodes with high data quality and large sample size to play a greater role, improve the performance of the global model, and iterates training until the loss function converges, ensuring the stability of the model and diagnostic accuracy.

[0019] Multiple specialist doctors perform secondary annotation, combined with weighted Kappa coefficient and Dice similarity coefficient for dual evaluation, and supplemented by a review process by doctors with the title of associate chief physician or above. This completely solves the problem of inconsistent annotation standards among different doctors and provides the model with highly reliable annotation data. The feedback loop uses the review data to fine-tune the local model, continuously optimize the lesion identification weight, and improve the model's ability to accurately identify lesions. Attached Figure Description

[0020] Figure 1 This is a flowchart of a deep learning-based intelligent diagnostic method for medical images. Figure 2 This is an internal framework diagram of a deep learning-based intelligent medical imaging diagnostic system. Detailed Implementation

[0021] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0022] Reference Figure 1 As shown, a deep learning-based intelligent diagnostic method for medical images includes: It receives medical image data from multiple independent medical institutions, with each institution's image data accompanied by an initial diagnostic label; it performs privacy protection processing on the image data from each medical institution; and it constructs an adaptive data augmentation module based on anatomical structure to generate augmented samples that conform to the anatomical characteristics of the human organ corresponding to the image. A hierarchical federated learning framework is constructed, consisting of a central server node and multiple edge nodes in medical institutions. Each edge node deploys an initial deep learning model with the same structure. Each edge node trains the model using locally preprocessed image data. The central server node performs weighted aggregation of the gradients uploaded by each edge node, generates global model parameters after aggregation, and feeds them back to each edge node. The training is iterated until the model loss function converges. A multi-dimensional annotation consistency assessment module was constructed, in which multiple specialists with certain clinical experience performed secondary annotation on the same batch of image data. The annotation content included the boundary coordinates of the lesion area and the classification of the lesion nature. Annotation consistency index was calculated, including the weighted Kappa coefficient and the Dice similarity coefficient. If any index was lower than the preset threshold, the annotation review process was initiated, and the final judgment was made by a doctor with the title of associate chief physician or above. A closed-loop annotation-model feedback mechanism is established, using the reviewed and annotated data as a supplementary training set, which is then input into the edge nodes of the federated learning framework to fine-tune the local model and update the lesion recognition weights. Medical image data to be diagnosed is received and input into the trained global deep learning model. The model outputs a heatmap of the lesion region, the probability distribution of lesion characteristics, and diagnostic confidence. If the diagnostic confidence is lower than a predetermined proportion, a manual review reminder is automatically triggered, and a report on the model's analysis of lesion characteristics is output.

[0023] It should be noted that differential privacy over homomorphic encryption is used in the gradient upload stage to ensure that the original data does not leave the domain and to prevent model inversion attacks from reconstructing patient images. The central server dynamically adjusts the aggregation weights based on the data quality scores of each institution, so that institutions with high-quality data have greater say and the robustness of the overall model is improved.

[0024] Organ partitioning involves constructing an atlas library containing three-dimensional morphological templates of human organs. During enhancement, deformation field constraints are used to ensure that the generated lesion morphology conforms to the actual anatomical spatial distribution. Pathological-physiological correlation enhancement is also performed. Enhancement strategies are customized for the lesion characteristics of different organs to ensure that the synthesized data has both morphological and biological rationality.

[0025] To ensure consistency in annotation, a multi-level arbitration mechanism is implemented. Initial annotations are completed independently by three attending physicians. When the weighted Kappa coefficient is less than 0.8, heatmap-assisted arbitration is triggered. The disputed area is visualized using Grad-CAM and submitted to the associate chief physician for final arbitration based on dynamic 3D reconstructed images. The arbitration results automatically update the annotation rule base, forming a continuously evolving annotation knowledge graph. The model is fine-tuned based on feedback, and the verified labeled data is injected into the local training set in the form of adversarial examples. The model's ability to identify easily confused lesions is enhanced by FocalLoss weighting.

[0026] For data-scarce institutions, a federated enhanced sample pool is provided. Anatomical-compliant enhanced data generated by each edge node is de-identified and then merged into the central cross-domain sample pool to support rare disease institutions in accessing synthetic data. For institutions without cases, a cross-modal generation model (Text-to-CT) based on pathological text description is provided, which generates synthetic images that conform to anatomy by inputting disease feature text.

[0027] The privacy protection processing of image data from various medical institutions specifically includes: Differential privacy technology is used to inject noise of a preset strength into the identity field metadata containing the patient's name and medical record number; homomorphic encryption algorithm is used to encrypt the key feature data of the image's pixel features and texture features, and the encrypted data can be used in subsequent model training without decryption.

[0028] The construction of the adaptive data augmentation module based on anatomical structure, which generates augmented samples that conform to the anatomical features of the human organ corresponding to the image, specifically includes: An anatomical database is established to structurally store information on the morphological characteristics, tissue density distribution patterns, and common lesion areas of different organs; the human organ partitions corresponding to the input images are automatically identified through image segmentation algorithms. Conventional enhanced sample generation scales the organ region to its original size based on organ boundary coordinates to simulate the image representation of changes in imaging distance or organ morphology; and adjusts the image grayscale values ​​by a predetermined ratio with reference to the grayscale range of normal and diseased tissues to simulate the grayscale differences of scanning parameters and pathological conditions. If a single institution has insufficient rare disease image samples, the generative adversarial network submodule is activated. Based on existing rare disease images, it learns the association pattern between lesions and background, and generates simulated samples with differences in core lesion features less than or equal to a predetermined proportion to supplement rare disease data.

[0029] It should be noted that differential privacy employs a dynamic noise injection mechanism. For identity field metadata (name, medical record number), a Laplace noise model Lap(Δf / ϵ) is used, where the sensitivity Δf is dynamically adjusted based on the field correlation. The name field has a Δf = 1.5. =0.8 (high-strength protection); The medical record number field has a value of Δf=0.7. =1.2 (Balanced Availability); By quantifying the strength of the association between fields and patient identity, we can achieve differentiated allocation of privacy budgets and avoid data utility collapse caused by uniform noise. Metadata perturbation verification involves injecting noise and then performing a k-anonymity test (k≥5) to ensure that any record cannot be associated with fewer than 5 individuals, thus meeting the GDPR anonymization standard.

[0030] Homomorphic encryption: It adopts the Paillier semi-homomorphic encryption algorithm, which supports feature value addition and scalar multiplication in ciphertext state, meeting the needs of image feature aggregation; Key fragment management: each medical institution holds a fragment of its private key, and the central server requires at least 3 institutions to jointly decrypt the data, thus preventing single-point data leakage from the source.

[0031] Anatomical structure database:

[0032] Morphological enhancement, scaling operations introduce organ deformation constraint functions:

[0033] In the formula, This represents the maximum permissible value for organ deformation. This refers to the standard organ area, i.e., the reference area. is the actual organ area; k is the deformation threshold, which is given as 0.35 here; To prevent the generation of images that violate anatomical rules (such as kidneys stretched to 130% beyond their original size).

[0034] Grayscale simulation, dynamically adjusting window width and level based on organ pathological state: During the inflammatory phase, the CT value of the liver parenchyma increases by 20-40 HU; during the fibrotic phase, it decreases by 15-25 HU and increases the gray variance. This avoids the misdiagnosis of fatty liver as normal liver due to traditional enhancement methods.

[0035] Generative Adversarial Networks (GANs) are pathological feature decouplers that decompose the output of generator G into:

[0036] In the formula, The input signal for synthesis; This is a function of the generator, where z typically represents the time-shift operator for discrete-time signals; Background organs, that is, the normal characteristics and signals of organs when there is no disease; The diseased component refers to the part of the organ whose characteristics and signals have changed due to the lesion; constraining the diseased component. The core feature difference δ≤5%.

[0037] The construction of the hierarchical federated learning framework specifically includes: The central server node is configured with a parameter aggregation engine, a global model storage module, and node communication authentication. It is responsible for receiving model gradients uploaded by edge nodes, performing aggregation calculations, generating a global model, and distributing it to each edge node. At the same time, it records the model update logs of each node through blockchain technology. The edge nodes of the medical institution are each equipped with a local data storage unit, a model training engine and an encrypted communication module; each edge node is pre-deployed with an initial deep learning model with a consistent structure. The model input layer is adapted to the medical image modality, the hidden layer contains 3-5 residual blocks, and the output layer is a lesion classification probability vector.

[0038] The model training at each edge node using locally preprocessed image data specifically includes: Edge nodes read preprocessed image data from local storage units and divide it into training and validation sets according to a predetermined ratio. The training set is grouped by organ type, and each group uses class-balanced sampling. Local training parameters are trained using the Adam optimizer, with batch size adjusted according to node computing power, and weighted cross-entropy is used as the loss function. After each round of local training is completed, the edge nodes calculate the gradient of the model parameters, encrypt the gradient using an asymmetric encryption algorithm, and transmit the encrypted gradient data only to the central server. The original images and intermediate training data are always stored locally.

[0039] The central server node performs weighted aggregation of gradients uploaded from each edge node, generates global model parameters after aggregation, and feeds them back to each edge node. Iterative training until the model loss function converges specifically includes: Gradient-weighted aggregation: After the central server decrypts the gradients of each edge node, it calculates the aggregation weights according to the formula; by calculating the cosine similarity between the gradient vector and the historical average gradient, if the similarity is less than a predetermined threshold, it is judged as an abnormal gradient and corrected by the truncated mean method. After aggregation, global model parameters are generated. After model compression, the parameters are encrypted and distributed to each edge node. The edge node updates its local model with the global parameters and repeats the local training, gradient uploading, and parameter update process until the loss function of the validation set decreases by less than a predetermined threshold for three consecutive rounds.

[0040] It should be noted that the central server uses a blockchain auditing mechanism, generating an immutable hash fingerprint for each gradient upload and recording: (node ​​ID, gradient hash value, timestamp, aggregate weight); it adopts a medical-specific consortium blockchain, which is only authorized to hospitals and drug regulatory authorities, meeting the FDA 21 CFR Part 11 electronic record compliance requirements. Dynamic communication authentication and node authentication integrate X.509 certificates and physician ID card biometric dual-factor verification to prevent unauthorized nodes from stealing the model.

[0041] Edge nodes and residual blocks introduce a lesion-sensing attention mechanism:

[0042] In the formula, This is the output of the attention mechanism; Image features are those extracted from the input image. This is for pre-extracting lesion templates, i.e., feature templates related to the lesion; For activation functions; This is the weight matrix, used to map the input features to the attention space; For computationally adaptive training, for structures with GPU memory <8GB: gradient accumulation is automatically enabled (batch_size=16 → virtual batch_size=64); for CPU-only structures, TensorRT engine quantization is used, improving inference speed by 4.1 times.

[0043] Global aggregation, gradient weighting, aggregation weight formula:

[0044] In the formula, Let be the aggregate weight of the i-th institution; Let i be the amount of data for the i-th institution; This represents the total amount of data across all institutions. The data quality score for the i-th institution is calculated based on factors such as annotation consistency and image signal-to-noise ratio. Historical gradient similarity is used to measure the similarity of historical data. When this value is greater than 0.85, it indicates a high similarity, and the data can be fully adopted. Abnormal gradient medical correction: the mean truncation method retains the middle 60% of gradient values. Triggering condition: cos(θ) < 0.6 and two consecutive abnormalities → automatically initiate institutional data quality review.

[0045] The calculated consistency index for annotation specifically includes: The weighted Kappa coefficient is calculated to assess the consistency of lesion classification. A contingency table is constructed based on the doctor's annotation results, and a clinical weight factor is introduced. The classification consistency is quantified by calculating the ratio of the actual consistency rate to the expected consistency rate. The Dice similarity coefficient calculation is based on the consistency of lesion boundaries. It extracts lesion areas labeled by different doctors, calculates the overlap by the ratio of twice the area intersection to the sum of the area pixels, and sets weights according to the difference in importance between the core area and the edge area of ​​the lesion to obtain the comprehensive Dice coefficient. Based on the accuracy requirements of medical diagnosis and industry consensus, preset qualification thresholds for two types of indicators were determined as the criteria for judging whether the labeling consistency meets the standards.

[0046] It should be noted that the weighted Kappa coefficient is: Medical weighting factors, misdiagnosis cost matrix design, weighting factors defined according to WHO clinical guidelines. :

[0047] Actual Consistency Rate (PoPo) Calculation:

[0048] In the formula, This represents the actual consistency rate. The weights are from category i to category j; N represents the number of samples in which physician A is classified as class i and physician B is classified as class j; N is the total number of samples. This represents the maximum weight, given here as 3.5, used for the conversion from malignant to benign. Grading determination strategy: κ≥0.85: Excellent consistency (can be directly used for model training); 0.75≤κ<0.85: Initiate automatic dispute detection; κ<0.75: Forced manual arbitration (trigger rate ≤5% is considered as system compliance).

[0049] Special disease adjustment: Pancreatic cancer, threshold increased to 0.90 (strong pathological heterogeneity). Lung nodules, threshold dropped to 0.80 (highly affected by imaging slice thickness).

[0050] Improved Dice formula:

[0051] In the formula, W is the weighted Dice coefficient, used to measure the similarity between two labeled regions; W is the pixel-level weight matrix, used to assign different weights to different pixels; A and B are two different physician labeled regions, or the region predicted by the model and the real labeled region. To show the intersection of regions A and B; Clinical guidelines for handling boundary disputes:

[0052] The establishment of the annotation-model feedback closed loop specifically includes: Supplementary training set processing involves dividing the final labeled data after verification into a supplementary training subset and a validation subset according to a certain ratio, and associating them with corresponding image feature labels. The local model fine-tuning mechanism involves the federated learning edge nodes receiving supplementary training sets and then using mini-batch gradient descent to fine-tune the local model, reducing the learning rate to 1 / 5 of the initial training rate. The focus is on updating the weights of convolutional layers related to lesion identification in the model, and setting a 1.5-fold learning weight for the lesion features corrected in the verification annotation. After weight update verification, the model performance is evaluated by supplementing the verification subset after fine-tuning. If the lesion recognition accuracy improves by more than or equal to a predetermined proportion, the updated local model weights are retained and participate in the next round of federated aggregation; otherwise, the weights are backtracked to the weights before fine-tuning, and the difference features are recorded for optimization in the next supplementary training.

[0053] The model outputs a heatmap of the lesion area, a probability distribution of the lesion's nature, and a diagnostic confidence level, specifically including: Upon receiving the medical images to be diagnosed, the system automatically performs privacy-preserving and standardization processing on the same data source as the training data to ensure it matches the model's input requirements. The pre-processed images are then input into the trained global deep learning model. The model uses a feature extraction layer to locate potential lesion areas and a classification layer to output multi-dimensional results. The heatmap of the lesion area highlights high-probability lesion areas identified by the model with a color gradient, and the overlay of coordinate grid facilitates localization; the probability distribution of lesion nature outputs the predicted probabilities of benign, malignant, and suspected lesions in percentage form, with the highest probability item being the preliminary judgment result; it is calculated by combining the lesion feature matching degree and the model output probability entropy value, with the lower the entropy value, the higher the confidence level.

[0054] It should be noted that the supplementary training set is processed in a hierarchical manner, with a data stratification strategy that prioritizes disputed samples and allocates arbitration-corrected samples to the supplementary training subset (accounting for ≥70%) to enhance the model's recognition of ambiguous lesions; rare disease samples are protected, with all rare cases entering the training subset, triggering 2x oversampling + adversarial enhancement (FGSM perturbation). Enhanced label association: In addition to lesion classification labels, radiomics labels (128-dimensional features such as texture and morphology) are simultaneously associated.

[0055] Dynamic performance threshold: For routine lesions: accuracy improvement ≥3% (meeting the minimum clinical benefit criteria for RSNA); Malignant lesions: Sensitivity increased by ≥5% (to prevent missed diagnoses); Cognitive mining through retrospective analysis: differential feature clustering → identifying high-frequency misjudgment patterns (such as misjudging inflammatory pseudotumors as liver cancer); generating counterfactual samples and injecting them into the next round of training.

[0056] Model output: Anatomical intelligence of lesion heatmaps, spatial localization enhancement technology, heatmap generation and fusion gradient-weighted class activation map (Grad-CAM++) and attention mechanism; Clinically-friendly visualization: The coordinate grid adopts the RAS medical coordinate system (the standard for neurosurgery) and supports seamless integration with PACS systems; Adjustable thermal threshold: Physicians can drag the slider to switch the confidence level display range (30%-99%) in real time. Decision support for lesion nature probability: Probability calibration technique, using temperature scaling to calibrate the output, to solve the problem of model overconfidence, such as turning an actual malignancy probability of 60% into an output of 85%; Clinical grading logic:

[0057] Reference Figure 2 As shown, a deep learning-based intelligent medical imaging diagnostic system includes: Data acquisition and preprocessing module: used to acquire medical image data from multiple institutions, perform format standardization, noise removal and equipment deviation correction on the image data, and output preprocessed image data; A hierarchical federated learning framework consists of a central server node, which includes a parameter aggregation engine, a global model storage unit, and node communication authentication; and multiple medical institution edge nodes, each of which deploys a local model training engine, an encrypted communication unit, and a local data storage unit. The local model training engine is pre-configured with an initial deep learning model of the same structure. The annotation consistency assessment module includes: an annotation task allocation unit, which selects specialists with clinical experience to perform secondary annotation on the same batch of image data, including the boundary coordinates of lesion areas and the classification of lesion nature; a consistency calculation unit, which calculates the weighted Kappa coefficient and Dice similarity coefficient; and a review trigger unit, which initiates the annotation review process when the weighted Kappa coefficient and Dice similarity coefficient are lower than the predetermined threshold, calling the terminal devices of doctors with the title of associate chief physician or above to make the final judgment. The annotation-model feedback closed-loop module includes: a supplementary training set construction unit, which divides the verified annotation data into a supplementary training subset and a validation subset; and a local model fine-tuning unit, which controls the edge nodes to fine-tune the local model using mini-batch gradient descent, focusing on updating the weights of the convolutional layers related to lesion recognition. Intelligent diagnostic reasoning module: Image input unit, which receives medical image data to be diagnosed and performs privacy protection processing; Model reasoning unit, which inputs the preprocessed image into the trained global model and outputs a heat map of the lesion area, probability distribution of lesion nature, and diagnostic confidence level; Manual review triggering unit, which automatically pushes a review reminder to the doctor's terminal when the diagnostic confidence level is lower than a predetermined threshold, and outputs an auxiliary report containing lesion feature analysis and reference cases.

[0058] It should be noted that the anatomical constraints for equipment deviation correction employ organ deformation field mapping technology, constructing a non-rigid registration model based on the anatomical boundary points of organs such as the liver / lung (e.g., Couinaud liver segmentation markers).

[0059] In the formula, To minimize the deformation field T; The deformation field T acts on the source image point Transformation on; The corresponding point in the reference image; is the regularization parameter used to balance the data fidelity term and the smoothing term; is the smoothing constraint term for the deformation field T to ensure the smoothness of the deformation field; N is the total number of sample points; ensures that the spatial consistency error of images from different CT models is <0.5mm; The lesion protection mechanism of noise removal, in low-dose CT denoising, constrains the denoising process through the lesion perception loss function to avoid the false elimination of small nodules (<3mm).

[0060] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A deep learning-based intelligent diagnostic method for medical images, characterized in that, include: It receives medical image data from multiple independent medical institutions, with each institution's image data accompanied by an initial diagnostic label; it performs privacy protection processing on the image data from each medical institution; and it constructs an adaptive data augmentation module based on anatomical structure to generate augmented samples that conform to the anatomical characteristics of the human organ corresponding to the image. A hierarchical federated learning framework is constructed, consisting of a central server node and multiple edge nodes in medical institutions. Each edge node deploys an initial deep learning model with the same structure. Each edge node trains the model using locally preprocessed image data. The central server node performs weighted aggregation of the gradients uploaded by each edge node, generates global model parameters after aggregation, and feeds them back to each edge node. The training is iterated until the model loss function converges. A multi-dimensional annotation consistency assessment module was constructed, in which multiple specialists with certain clinical experience performed secondary annotation on the same batch of image data. The annotation content included the boundary coordinates of the lesion area and the classification of the lesion nature. Annotation consistency index was calculated, including the weighted Kappa coefficient and the Dice similarity coefficient. If any index was lower than the preset threshold, the annotation review process was initiated, and the final judgment was made by a doctor with the title of associate chief physician or above. A closed-loop annotation-model feedback mechanism is established, using the reviewed and annotated data as a supplementary training set, which is then input into the edge nodes of the federated learning framework to fine-tune the local model and update the lesion recognition weights. Medical image data to be diagnosed is received and input into the trained global deep learning model. The model outputs a heatmap of the lesion region, the probability distribution of lesion characteristics, and diagnostic confidence. If the diagnostic confidence is lower than a predetermined proportion, a manual review reminder is automatically triggered, and a report on the model's analysis of lesion characteristics is output.

2. The intelligent medical image diagnosis method based on deep learning according to claim 1, characterized in that, The privacy protection processing of image data from various medical institutions specifically includes: Differential privacy technology is used to inject noise of a preset strength into the identity field metadata containing the patient's name and medical record number; homomorphic encryption algorithm is used to encrypt the key feature data of the image's pixel features and texture features, and the encrypted data can be used in subsequent model training without decryption.

3. The intelligent medical image diagnosis method based on deep learning according to claim 2, characterized in that, The construction of the adaptive data augmentation module based on anatomical structure, which generates augmented samples that conform to the anatomical features of the human organ corresponding to the image, specifically includes: An anatomical database is established to structurally store information on the morphological characteristics, tissue density distribution patterns, and common lesion areas of different organs; the human organ partitions corresponding to the input images are automatically identified through image segmentation algorithms. Conventional enhanced sample generation scales the organ region to its original size based on organ boundary coordinates to simulate the image representation of changes in imaging distance or organ morphology; and adjusts the image grayscale values ​​by a predetermined ratio with reference to the grayscale range of normal and diseased tissues to simulate the grayscale differences of scanning parameters and pathological conditions. If a single institution has insufficient rare disease image samples, the generative adversarial network submodule is activated. Based on existing rare disease images, it learns the association pattern between lesions and background, and generates simulated samples with differences in core lesion features less than or equal to a predetermined proportion to supplement rare disease data.

4. The intelligent medical image diagnosis method based on deep learning according to claim 1, characterized in that, The construction of the hierarchical federated learning framework specifically includes: The central server node is configured with a parameter aggregation engine, a global model storage module, and node communication authentication. It is responsible for receiving model gradients uploaded by edge nodes, performing aggregation calculations, generating a global model, and distributing it to each edge node. At the same time, it records the model update logs of each node through blockchain technology. The edge nodes of the medical institution are each equipped with a local data storage unit, a model training engine and an encrypted communication module; each edge node is pre-deployed with an initial deep learning model with a consistent structure. The model input layer is adapted to the medical image modality, the hidden layer contains 3-5 residual blocks, and the output layer is a lesion classification probability vector.

5. The intelligent medical image diagnosis method based on deep learning according to claim 4, characterized in that, The model training at each edge node using locally preprocessed image data specifically includes: Edge nodes read preprocessed image data from local storage units and divide it into training and validation sets according to a predetermined ratio. The training set is grouped by organ type, and each group uses class-balanced sampling. Local training parameters are trained using the Adam optimizer, with batch size adjusted according to node computing power, and weighted cross-entropy is used as the loss function. After each round of local training is completed, the edge nodes calculate the gradient of the model parameters, encrypt the gradient using an asymmetric encryption algorithm, and transmit the encrypted gradient data only to the central server. The original images and intermediate training data are always stored locally.

6. The intelligent medical image diagnosis method based on deep learning according to claim 5, characterized in that, The central server node performs weighted aggregation of gradients uploaded from each edge node, generates global model parameters after aggregation, and feeds them back to each edge node. Iterative training until the model loss function converges specifically includes: Gradient-weighted aggregation: After the central server decrypts the gradients of each edge node, it calculates the aggregation weights according to the formula; by calculating the cosine similarity between the gradient vector and the historical average gradient, if the similarity is less than a predetermined threshold, it is judged as an abnormal gradient and corrected by the truncated mean method. After aggregation, global model parameters are generated. After model compression, the parameters are encrypted and distributed to each edge node. The edge node updates its local model with the global parameters and repeats the local training, gradient uploading, and parameter update process until the loss function of the validation set decreases by less than a predetermined threshold for three consecutive rounds.

7. The intelligent medical image diagnosis method based on deep learning according to claim 1, characterized in that, The calculated consistency index for annotation specifically includes: The weighted Kappa coefficient is calculated to assess the consistency of lesion classification. A contingency table is constructed based on the doctor's annotation results, and a clinical weight factor is introduced. The classification consistency is quantified by calculating the ratio of the actual consistency rate to the expected consistency rate. The Dice similarity coefficient calculation is based on the consistency of lesion boundaries. It extracts lesion areas labeled by different doctors, calculates the overlap by the ratio of twice the area intersection to the sum of the area pixels, and sets weights according to the difference in importance between the core area and the edge area of ​​the lesion to obtain the comprehensive Dice coefficient. Based on the accuracy requirements of medical diagnosis and industry consensus, preset qualification thresholds for two types of indicators were determined as the criteria for judging whether the labeling consistency meets the standards.

8. The intelligent medical image diagnosis method based on deep learning according to claim 1, characterized in that, The establishment of the annotation-model feedback closed loop specifically includes: Supplementary training set processing involves dividing the final labeled data after verification into a supplementary training subset and a validation subset according to a certain ratio, and associating them with corresponding image feature labels. The local model fine-tuning mechanism involves the federated learning edge nodes receiving supplementary training sets and then using mini-batch gradient descent to fine-tune the local model, reducing the learning rate to 1 / 5 of the initial training rate. The focus is on updating the weights of convolutional layers related to lesion identification in the model, and setting a 1.5-fold learning weight for the lesion features corrected in the verification annotation. After weight update verification, the model performance is evaluated by supplementing the verification subset after fine-tuning. If the lesion recognition accuracy improves by more than or equal to a predetermined proportion, the updated local model weights are retained and participate in the next round of federated aggregation; otherwise, the weights are backtracked to the weights before fine-tuning, and the difference features are recorded for optimization in the next supplementary training.

9. The intelligent medical image diagnosis method based on deep learning according to claim 8, characterized in that, The model outputs a heatmap of the lesion area, a probability distribution of the lesion's nature, and a diagnostic confidence level, specifically including: Upon receiving the medical images to be diagnosed, the system automatically performs privacy-preserving and standardization processing on the same data source as the training data to ensure it matches the model's input requirements. The pre-processed images are then input into the trained global deep learning model. The model uses a feature extraction layer to locate potential lesion areas and a classification layer to output multi-dimensional results. The heatmap of the lesion area highlights high-probability lesion areas identified by the model with a color gradient, and the overlay of coordinate grid facilitates localization; the probability distribution of lesion nature outputs the predicted probabilities of benign, malignant, and suspected lesions in percentage form, with the highest probability item being the preliminary judgment result; it is calculated by combining the lesion feature matching degree and the model output probability entropy value, with the lower the entropy value, the higher the confidence level.

10. A deep learning-based intelligent medical imaging diagnostic system, characterized in that, include: Data acquisition and preprocessing module: used to acquire medical image data from multiple institutions, perform format standardization, noise removal and equipment deviation correction on the image data, and output preprocessed image data; A hierarchical federated learning framework: a central server node, including a parameter aggregation engine, a global model storage unit, and node communication authentication; Multiple medical institution edge nodes, each node deploys a local model training engine, an encrypted communication unit and a local data storage unit, and the local model training engine is pre-configured with an initial deep learning model with the same structure; The annotation consistency assessment module includes: an annotation task allocation unit, which selects specialists with clinical experience to perform secondary annotation on the same batch of image data, including the boundary coordinates of lesion areas and the classification of lesion nature; a consistency calculation unit, which calculates the weighted Kappa coefficient and Dice similarity coefficient; and a review trigger unit, which initiates the annotation review process when the weighted Kappa coefficient and Dice similarity coefficient are lower than the predetermined threshold, calling the terminal devices of doctors with the title of associate chief physician or above to make the final judgment. The annotation-model feedback closed-loop module includes: a supplementary training set construction unit, which divides the verified annotation data into a supplementary training subset and a validation subset; and a local model fine-tuning unit, which controls the edge nodes to fine-tune the local model using mini-batch gradient descent, focusing on updating the weights of the convolutional layers related to lesion recognition. Intelligent diagnostic reasoning module: Image input unit, which receives medical image data to be diagnosed and performs privacy protection processing; The model inference unit inputs the preprocessed images into the trained global model and outputs a heat map of the lesion area, a probability distribution of the lesion nature, and a diagnostic confidence level. The manual review triggering unit automatically pushes a review reminder to the doctor's terminal when the diagnostic confidence level is lower than a predetermined threshold, and outputs an auxiliary report containing lesion feature analysis and reference cases.

Images