A vocal cord white lesion diagnosis method and system based on a pre-trained large model and attention transfer

Abstract

The application discloses a vocal cord white lesion diagnosis method and system based on a pre-training large model and attention transfer, and relates to the technical field of medical image processing and artificial intelligence auxiliary diagnosis. Frames are extracted from a laryngoscope video, a standardized image dataset is constructed, and a pre-trained visual backbone network is used to extract multi-scale features, and an exogenous probability heat map is generated as visual prior by means of a multi-modal medical basic model. By calibrating the prior to the attention domain and aligning it with the internal attention of the network, the model is guided to focus on the lesion area. A two-stage classification strategy is adopted, first screening "cancer" and "non-cancer", and then subdividing "non-cancer" lesions into "special infection or inflammation", "vocal cord white spot-low risk" and "vocal cord white spot-high risk", and improving robustness through ensemble learning. The system synchronously outputs an interpretable heat map and a structured diagnosis conclusion, and supports clinical interaction. It effectively improves the consistency of diagnosis, clinical applicability and work efficiency.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence-assisted diagnosis, and in particular to a diagnostic method and system for vocal cord leukoplakia based on a pre-trained large model and attention transfer. Background Technology

[0002] Vocal cord leukoplakia is a common clinical finding in the larynx, encompassing various benign and malignant lesions such as vocal cord leukoplakia, keratosis, specific vocal cord infections, and early squamous cell carcinoma. Early diagnosis and accurate grading are crucial for selecting clinical treatment plans and patient prognosis. Currently, the clinical diagnosis of vocal cord leukoplakia primarily relies on the visual evaluation of electronic laryngoscopy images by physicians, supplemented by enhancement modalities such as narrow-band imaging (NBI). Experienced physicians can differentiate between benign and malignant lesions based on their morphology, boundaries, and vascular patterns; however, this method is subject to significant subjective variability, especially in early lesions, small lesions, or those with indistinct borders, potentially leading to missed or misdiagnosis.

[0003] Pathological tissue biopsy is currently the gold standard for diagnosing the nature of vocal cord lesions. However, it is an invasive procedure that may cause complications such as vocal cord injury, bleeding, and infection, and it carries the risk of sampling errors, especially for multifocal or diffuse lesions where comprehensive coverage is difficult. Therefore, there is an urgent clinical need for a non-invasive, repeatable, objective, and accurate auxiliary diagnostic method.

[0004] In recent years, with the rapid development of artificial intelligence technology in the field of medical image analysis, some studies have attempted to automatically classify and detect lesions in laryngoscope images based on deep learning models. Existing methods mostly focus on binary classification tasks (such as benign versus malignant), employing end-to-end training architectures such as convolutional neural networks (CNNs) or visual transformers (ViTs). However, these methods have the following limitations: Insufficient fine-grained stratification capability: In clinical practice, vocal cord leukoplakia needs to be further divided into subclasses such as low-risk leukoplakia, high-risk leukoplakia, inflammatory or infectious lesions. Existing models generally lack the ability to perform fine-grained risk stratification for non-cancerous lesions, making it difficult to meet the needs of precise clinical management.

[0005] Weak interpretability: Most deep learning models operate as "black boxes," failing to provide convincing visual explanations. Doctors struggle to understand the model's decision-making rationale, reducing its acceptance and credibility in clinical practice.

[0006] Reliance on fine annotation: Some studies have attempted to introduce lesion segmentation tasks to improve localization capabilities, but they rely on pixel-level fine annotation, which is costly, time-consuming, and difficult to guarantee in terms of annotation consistency, thus limiting the scalable application and generalization performance of the model.

[0007] Disconnected from clinical workflows: Existing methods mostly focus on offline analysis and do not fully consider seamless integration with hospital endoscopy workstations, electronic medical record systems, etc., nor do they effectively fuse and utilize multimodal images (such as white light and NBI images).

[0008] Long-tail data problem: The scarcity of samples in categories such as high-risk lesions and special infections leads to unstable recognition performance of the model in a few categories, affecting the overall balance and robustness of diagnosis.

[0009] Therefore, how to construct an intelligent auxiliary system capable of fine-grained hierarchical diagnosis of vocal cord leukoplakia without relying on fine annotation, possessing good interpretability, and easily integrating with clinical workflows, has become a key technical problem urgently needing to be solved in this field. This invention is proposed against this backdrop, aiming to overcome the limitations of existing methods and improve the accuracy, reliability, and clinical applicability of intelligent diagnosis of vocal cord lesions by introducing prior knowledge from a pre-trained large model and an attention transfer mechanism. Summary of the Invention

[0010] The purpose of this invention is to overcome the aforementioned shortcomings of existing technologies and to provide a diagnostic method and system for vocal cord leukoplakia based on a pre-trained large model and attention transfer. By introducing extrinsic visual priors and a hierarchical attention alignment mechanism, stable localization of the lesion area and fine-grained risk stratification are achieved, and interpretable heatmaps are output simultaneously to improve the objectivity, accuracy, and clinical reliability of the diagnosis, assisting doctors in achieving early screening and accurate decision-making.

[0011] To achieve the above-mentioned objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a diagnostic method for vocal cord leukoplakia based on a pre-trained large model and attention transfer, comprising the following steps: S1. Data Processing: Acquire laryngoscopy videos, perform frame extraction, de-identification, quality screening, and field of view standardization to construct a standardized image dataset; S2. Feature Extraction: Extract multi-scale visual features from the input image using a pre-trained visual backbone network; S3, Prior Generation: Using a frozen multimodal medical vision basic model, reasoning is performed on the input image to generate an external probability heatmap P corresponding to the spatial location of the input image; S4. Prior transformation: The external probability heatmap P is calibrated through a lightweight convolutional network and mapped to the attention domain using conditional modulation to obtain the attention prior Q. S5. Attention Alignment: Align the attention prior Q with the student attention distribution generated by the visual backbone network at multiple scales to guide the network to focus on the lesion-related area. S6. First-stage classification: Under the alignment constraint, a binary classifier is trained to perform initial screening classification of the input image as "cancer" or "non-cancer", and outputs the first classification probability and the corresponding first interpretable heatmap. S7. Second-stage classification: For image samples classified as "non-cancer", multiple isomorphic learners are initialized. Under the premise of freezing the parameters of the prior generation and attention alignment module, only the visual backbone network and classification head are fine-tuned to perform three classifications: "special infection or inflammation", "vocal cord leukoplakia - low risk" and "vocal cord leukoplakia - high risk". The outputs of each learner are integrated to obtain the second classification probability. S8. Result Output and Visualization: The final diagnostic conclusion is generated by combining the first classification probability and the second classification probability, and the corrected interpretable heatmap is superimposed on the diagnostic conclusion for visualization.

[0012] Furthermore, the process of mapping the external probability heatmap P to the attention prior Q in step S4 is expressed as follows: ; in, This refers to the lightweight convolutional network. Its parameters, This represents the conditional modulation function. is the modulation coefficient.

[0013] Furthermore, the attention alignment described in step S5 is achieved by calculating the attention prior Q and the student attention distribution at each scale. The KL divergence loss between them is realized, the first Layer alignment loss Represented as: ; in, In order to be with the first The attention prior for layer spatial resolution matching is used, and the total alignment loss is the sum of the losses of each layer.

[0014] Furthermore, in step S6, the total loss function of the first stage classification Cross-entropy loss for binary classification Compared with the total alignment loss described in step S5 Weighted sum: ; in, To balance the hyperparameters of the two losses.

[0015] Furthermore, the integration of the outputs of multiple isomorphic learners described in step S7 specifically includes: firstly, temperature calibration of the output logic values ​​of each learner, and then weighted averaging or voting on the calibrated probabilities to obtain the final second classification probability.

[0016] Furthermore, the temperature calibration employs learnable temperature parameters. For the output logical value of a single learner Its calibrated probability for: .

[0017] Furthermore, the pre-trained visual backbone network is an EfficientNet-V2, ConvNeXt, or SwinTransformer architecture; the multimodal medical vision foundation model is a publicly available model pre-trained on a large medical image-text pair dataset.

[0018] On the other hand, the present invention provides a diagnostic system for vocal cord leukoplakia to implement the above method, comprising: The data processing and management module is used to execute step S1 in the above method; The image encoding module is used to perform step S2 in the above method; The prior extraction module is used to perform step S3 in the above method; The prior calibration module is used to perform step S4 in the above method; The hierarchical attention alignment module is used to perform step S5 in the above method; The first-stage classification module is used to execute step S6 in the above method; The second-stage segmentation and integration module is used to perform step S7 in the above method; An interpretable display and interaction module is used to perform step S8 in the above method.

[0019] Furthermore, it also includes a deployment and interface module, which is used to deploy the system to the hospital endoscopy workstation or server in the form of software service or hardware integration, and to interface with the hospital information system for data exchange.

[0020] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for diagnosing vocal cord leukoplakia.

[0021] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. Achieved non-invasive and accurate diagnosis and stratification: Through a two-stage "cancer / non-cancer" initial screening and a three-level subdivision of "non-cancer", accurate diagnosis and risk stratification of vocal cord leukoplakia can be achieved without invasive biopsy and pixel-level fine annotation, significantly improving the objectivity of diagnosis and clinical applicability.

[0022] 2. Improved model interpretability and clinical acceptance: The model utilizes a multimodal large model to generate exogenous priors and guides the model to focus on the lesion area through an attention alignment mechanism. While outputting classification results, it simultaneously generates a visualized and interpretable heatmap, making the model's decision-making process transparent to doctors and greatly enhancing clinical trust and system usability.

[0023] 3. Enhanced robustness in small sample sizes and complex scenarios: Through attention prior guidance and multi-learner ensemble strategies, the model training challenges caused by long-tail data distribution (such as the scarcity of high-risk vitiligo samples) and neighbor class confusion (such as inflammation and low-risk vitiligo) are effectively alleviated, improving the system's stability and generalization ability under conditions of imbalanced data and ambiguous boundaries.

[0024] 4. Reduced deployment costs and improved system efficiency: The "freeze prior - fine-tune backbone" training strategy and modular design greatly reduce training computation overhead; the inference phase only requires image forward propagation, meeting the real-time requirements of clinical practice; the system can be flexibly deployed to in-hospital workstations or the cloud, and can be seamlessly integrated with existing medical information systems, making it easy to promote and apply in clinical practice.

[0025] 5. A complete closed-loop auxiliary diagnostic workflow has been formed: from raw video frame extraction and processing, intelligent analysis and diagnosis, to structured report generation and visual interaction, this invention has constructed an end-to-end auxiliary diagnostic solution, which is effectively integrated into the actual clinical workflow, supports diagnostic record keeping and review, and provides reliable technical support for the standardized diagnosis and management of vocal cord lesions. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall structure of the system of the present invention; wherein (a) is the BiomedParse prior generation module, (b) is the binary classifier and feature extraction backbone, (c) is the details of the prior calibration and attention domain mapping module, and (d) is the integrated discriminant network architecture; Figure 2 This is a schematic diagram of a process of the method of the present invention; Figure 3 This is another schematic diagram of the method of the present invention; Figure 4 This is a schematic diagram of the attention alignment module structure in this invention; Figure 5 This is a schematic diagram illustrating the clinical validation of a diagnostic method for vocal cord leukoplakia based on a pre-trained large model and attention transfer. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0028] Example 1 This embodiment provides a diagnostic method for vocal cord leukoplakia based on a pre-trained large model and attention transfer, including the following steps: S1. Data Processing: Acquire laryngoscopy videos, perform frame extraction, de-identification, quality screening, and field of view standardization to construct a standardized image dataset; S2. Feature Extraction: Extract multi-scale visual features from the input image using a pre-trained visual backbone network; S3, Prior Generation: Using a frozen multimodal medical vision basic model, reasoning is performed on the input image to generate an external probability heatmap P corresponding to the spatial location of the input image; S4. Prior transformation: The external probability heatmap P is calibrated through a lightweight convolutional network and mapped to the attention domain using conditional modulation to obtain the attention prior Q. S5. Attention Alignment: Align the attention prior Q with the student attention distribution generated by the visual backbone network at multiple scales to guide the network to focus on the lesion-related area. S6. First-stage classification: Under the alignment constraint, a binary classifier is trained to perform initial screening classification of the input image as "cancer" or "non-cancer", and outputs the first classification probability and the corresponding first interpretable heatmap. S7. Second-stage classification: For image samples classified as "non-cancer", multiple isomorphic learners are initialized. Under the premise of freezing the parameters of the prior generation and attention alignment module, only the visual backbone network and classification head are fine-tuned to perform three classifications: "special infection or inflammation", "vocal cord leukoplakia - low risk" and "vocal cord leukoplakia - high risk". The outputs of each learner are integrated to obtain the second classification probability. S8. Result Output and Visualization: The final diagnostic conclusion is generated by combining the first classification probability and the second classification probability, and the corrected interpretable heatmap is superimposed on the diagnostic conclusion for visualization.

[0029] Specifically, this invention was implemented in the laryngoscopy center of a top-tier hospital. The doctor collects video footage of the patient's laryngoscopy examination, and the system follows... Figure 2 and Figure 3The process shown is executed automatically. First, S1 (Data Processing): The system extracts keyframes from the video at a step size of 2 frames per second, removes patient personal information identifiers, and retains frames with a sharpness > 0.8 through quality screening. Next, S2 (Feature Extraction): Features are extracted using a pre-trained visual backbone network. Then, S3 (Prior Generation): The BiomedParse model is called to generate an exogenous probability heatmap P. Following this, S4 (Prior Transformation): Attention prior Q is obtained through convolution. Next, S5 (Attention Alignment): Q is aligned with the activation maps of each layer of the network. S6 (First-Stage Classification): The probability of "Non-cancerous" is output as 0.92. Since the result is non-cancerous, S7 (Second-Stage Classification) is executed: The probability of "Vocal Cord Leukoplakia - High Risk" is output as 0.87. Finally, S8 (Result Output): The system generates a diagnostic report containing an interpretable heatmap.

[0030] Example 2 The process of mapping the external probability heatmap P to the attention prior Q in step S4 is expressed as follows: ; in, This refers to the lightweight convolutional network. Its parameters, This represents the conditional modulation function. is the modulation coefficient.

[0031] like Figure 1 As shown, the specific implementation process is as follows: When the system processes laryngoscope images, BiomedParse generates a 256×256 exogenous probability heatmap P. P first enters the prior calibration (Prior module), through two layers of 3×3 convolution (parameters...). The system first performs feature reconstruction; then it enters the conditional modulation unit, where learnable coefficients γ are used for semantic enhancement; finally, through the attention domain mapping unit, the probability distribution features are mapped to an attention prior Q aligned with the backbone network. The entire process strictly follows the formula. .

[0032] Example 3 The attention alignment described in step S5 is achieved by calculating the attention prior Q and the student attention distribution at each scale. The KL divergence loss between them is realized, the first Layer alignment loss Represented as: ; in, In order to be with the first The attention prior for layer spatial resolution matching is used, and the total alignment loss is the sum of the losses of each layer.

[0033] Specifically, an attention alignment mechanism is implemented during the training phase. For 512×512 laryngoscope images, such as... Figure 4As shown, the image encoding module generates student attention distributions at three levels. The information is transmitted via paths 1, 2, and 3 respectively. The attention prior Q obtained from the prior calibration module, after interpolation adjustment, is input along paths 4, 5, and 6. The KL divergence loss for each layer is calculated: First layer... The value is 0.15; second layer The value is 0.12; third layer The value is 0.10. Total alignment loss. It is 0.37.

[0034] Figure 4 Alignment calculation units 7, 8, and 9 in the middle perform loss calculations for these three layers respectively.

[0035] Example 4 The total loss function of the first stage classification in step S6 Cross-entropy loss for binary classification Compared with the total alignment loss described in step S5 Weighted sum: ; in, To balance the hyperparameters of the two losses.

[0036] Specifically, the features extracted by the image encoding module are input to the stage-one classification module, while the alignment loss calculated by the hierarchical attention alignment module is also input to the stage-one classification module. For 32 training images, the classification cross-entropy loss... The calculated alignment loss is 1.23. It is 0.37. Set λ=0.1, according to the formula... The total loss is calculated to be 1.267. The loss calculation unit within the Stage 1 classification module executes this formula, and the optimizer updates the network parameters based on this total loss.

[0037] Example 5 The integration of the outputs of multiple isomorphic learners in step S7 specifically includes: first, temperature calibration of the output logic values ​​of each learner, and then weighted averaging or voting on the calibrated probabilities to obtain the final second classification probability.

[0038] Specifically, when an image is classified as "non-cancerous," it enters the Stage 2 subdivision and ensemble module. This module contains five isomorphic learners. After each learner outputs a logical value vector, the ensemble unit within the Stage 2 subdivision and ensemble module first performs temperature calibration, and then performs a weighted average. Let the output probabilities of each learner be... , Equal weight final probability The system thus determined it to be a "special infection or inflammation".

[0039] Example 6 The temperature calibration uses learnable temperature parameters. For the output logical value of a single learner Its calibrated probability for: .

[0040] Specifically, for the output logic value z=[2.1,1.5,0.8] of a certain learner, the temperature calibration submodule of the stage two subdivision and integration module performs calibration calculations. Initially T=1.0, after optimization T=1.5. According to the formula... calculate, .

[0041] Example 7 The pre-trained visual backbone network is an EfficientNet-V2, ConvNeXt, or SwinTransformer architecture; the multimodal medical vision foundation model is a publicly available model pre-trained on a large medical image-text pair dataset.

[0042] Specifically, the image encoding module uses EfficientNet-V2s, and the prior extraction module uses the MedCLIP model. In scenario B, the image encoding module is replaced with SwinTransformer-Base, and the prior extraction module is replaced with PMC-CLIP. Both configurations run normally, realizing the complete process from the data processing and management module to the deployment and interface module, proving that the architecture of this invention is compatible with multiple pre-trained models.

[0043] Example 8 This invention provides a diagnostic system for vocal cord leukoplakia to implement the above method, comprising: The data processing and management module is used to execute step S1 in the above method; The image encoding module is used to perform step S2 in the above method; The prior extraction module is used to perform step S3 in the above method; The prior calibration module is used to perform step S4 in the above method; The hierarchical attention alignment module is used to perform step S5 in the above method; The first-stage classification module is used to execute step S6 in the above method; The second-stage segmentation and integration module is used to perform step S7 in the above method; An interpretable display and interaction module is used to perform step S8 in the above method.

[0044] Specifically, in a hospital deployment, the data processing and management module is implemented using Python / OpenCV; the image encoding module is implemented using PyTorch; the prior extraction module calls BiomedParse; the prior calibration module implements a lightweight convolutional network; the hierarchical attention alignment module calculates the KL divergence loss; the first-stage classification module includes fully connected layers; the second-stage segmentation and ensemble module implements multi-learner ensemble; and the interpretable display and interaction module is based on Web development.

[0045] Example 9 It also includes a deployment and interface module, which is used to deploy the system to the hospital endoscopy workstation or server in the form of software service or hardware integration, and to interface with the hospital information system for data exchange.

[0046] Specifically, in the implementation at a certain medical center, the deployment and interface module implemented the following interfaces: 1) DICOM interface to receive images from PACS; 2) HL7 interface to exchange information with HIS; 3) RESTful API to provide services to workstations; 4) Database interface to store results to EMR.

[0047] Example 10 This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for diagnosing vocal cord leukoplakia.

[0048] Specifically, the invention system is implemented in software form and can be stored on a computer-readable storage medium. When the program on the medium is executed, all module functions are fully implemented. Implementation method A: Burn the software containing all module code to the solid-state drive of the medical device; Implementation method B: Package it into an installation package and store it on a USB flash drive; Implementation method C: Provide a cloud server image file. Regardless of the method, when the program is loaded and executed from the medium, the system runs according to the above architecture, executing sequentially from the data processing and management module to the deployment and interface module, realizing intelligent diagnosis of vocal cord lesions.

[0049] Example 11 To verify the clinical effectiveness of the pre-trained large model and attention-shifting diagnostic method and system for vocal cord leukoplakia described in this invention, a 6-month clinical trial was conducted in the ENT department of a tertiary hospital. The study included 500 consecutive patients with suspected vocal cord leukoplakia who underwent laryngoscopy and were approved by the ethics committee as the validation set. All cases had pathological biopsy as the gold standard control. Figure 5As shown in (a), in the clinically critical binary screening task of "cancer" and "non-cancer," the model of this invention achieved an excellent performance of AUC 0.905. Compared with the comparative experiments conducted at the same time, this result is superior to the experienced team of expert physicians and the traditional AI baseline model without prior knowledge (AUC 0.82), demonstrating the significant role of introducing exogenous prior knowledge in improving the sensitivity of malignant tumor screening. For the clinically challenging subdivision of non-cancer lesions, such as... Figure 5 As shown in (b), the model exhibits extremely strong discriminative ability. It demonstrates exceptional accuracy in identifying distinctive "special infections" (AUC 0.947) and "squamous cell carcinoma" (AUC 0.905); it also maintains good discriminative power in morphologically easily confused "low-risk / inflammation" (AUC 0.821) and "high-risk leukoplakia" (AUC 0.778), with a macro-average AUC reaching 0.863. This indicates that the model effectively overcomes the class imbalance problem caused by the long-tailed distribution. Figure 5 As shown in (c), after employing an ensemble discriminant network, the overall multi-class classification performance was further stabilized at AUC 0.874. Furthermore, in clinical consistency assessment, the diagnostic results output by this invention showed a high degree of consistency with the pathological gold standard, with a Kappa coefficient reaching 0.82 (high consistency), significantly higher than the traditional single model, demonstrating extremely high clinical reference value. Figure 5 As shown in (e), the heatmap generated by this invention can accurately focus on the lesion and its irregular boundaries, which is significantly better than the diffuse heatmaps of traditional methods. This "what you see is what you get" visualization capability not only verifies that the model's decision-making is based on the lesion itself rather than background noise, but also effectively enhances doctors' trust in AI-assisted diagnosis.

[0050] This embodiment demonstrates through systematic clinical verification that the method of the present invention is not only superior to existing methods in terms of technical performance, but also has excellent interpretability, computational efficiency, and physician acceptance in actual clinical applications, providing reliable technical support for the intelligent auxiliary diagnosis of vocal cord leukoplakia.

[0051] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A diagnostic method for vocal cord leukoplakia based on a pre-trained large model and attention transfer, characterized in that, Includes the following steps: S1. Data Processing: Acquire laryngoscopy videos, perform frame extraction, de-identification, quality screening, and field of view standardization to construct a standardized image dataset; S2. Feature Extraction: Extract multi-scale visual features from the input image using a pre-trained visual backbone network; S3, Prior Generation: Using a frozen multimodal medical vision basic model, reasoning is performed on the input image to generate an external probability heatmap P corresponding to the spatial location of the input image; S4. Prior transformation: The external probability heatmap P is input into the prior calibration module. First, prior calibration is performed through a lightweight convolutional network. Then, semantic features are injected using conditional modulation. Finally, the distribution features are adjusted through attention domain mapping to obtain the attention prior Q that is aligned with the feature space of the visual backbone network. S5. Attention Alignment: Align the attention prior Q with the student attention distribution generated by the visual backbone network at multiple scales to guide the network to focus on the lesion-related area. S6. First-stage classification: Under the alignment constraint, a binary classifier is trained to perform initial screening classification of the input image as "cancer" or "non-cancer", and outputs the first classification probability and the corresponding first interpretable heatmap. S7. Second-stage classification: For image samples classified as "non-cancer", multiple isomorphic learners are initialized. Under the premise of freezing the parameters of the prior generation and attention alignment module, only the visual backbone network and classification head are fine-tuned to perform three classifications: "special infection or inflammation", "vocal cord leukoplakia - low risk" and "vocal cord leukoplakia - high risk". The outputs of each learner are integrated to obtain the second classification probability. S8. Result Output and Visualization: The final diagnostic conclusion is generated by combining the first classification probability and the second classification probability, and the corrected interpretable heatmap is superimposed on the diagnostic conclusion for visualization.

2. The method according to claim 1, characterized in that, The process of mapping the external probability heatmap P to the attention prior Q in step S4 is expressed as follows: ; in, This refers to the lightweight convolutional network. Its parameters, This represents the composite transform function of the conditional modulation function and the attention domain mapping. is the modulation coefficient.

3. The method according to claim 1 or 2, characterized in that, The attention alignment described in step S5 is achieved by calculating the attention prior Q and the student attention distribution at each scale. The KL divergence loss between them is realized, the first Layer alignment loss Represented as: ; in, In order to be with the first The attention prior for layer spatial resolution matching is used, and the total alignment loss is the sum of the losses of each layer.

4. The method according to claim 1, characterized in that, The total loss function of the first stage classification in step S6 Cross-entropy loss for binary classification Compared with the total alignment loss described in step S5 Weighted sum: ; in, To balance the hyperparameters of the two losses.

5. The method according to claim 1, characterized in that, The integration of the outputs of multiple isomorphic learners in step S7 specifically includes: first, temperature calibration of the output logic values ​​of each learner, and then weighted averaging or voting on the calibrated probabilities to obtain the final second classification probability.

6. The method according to claim 5, characterized in that, The temperature calibration uses learnable temperature parameters. For the output logical value of a single learner Its calibrated probability for: .

7. The method according to claim 1, characterized in that, The pre-trained visual backbone network is an EfficientNet-V2, ConvNeXt, or SwinTransformer architecture; the multimodal medical vision foundation model is a publicly available model pre-trained on a large medical image-text pair dataset.

8. A diagnostic system for vocal cord leukoplakia of any one of claims 1-7, characterized in that, include: The data processing and management module is used to perform step S1 in claim 1; The image encoding module is used to perform step S2 in claim 1; The prior extraction module is used to perform step S3 in claim 1; A priori calibration module is used to perform step S4 in claim 1; A hierarchical attention alignment module is used to perform step S5 in claim 1; The first-stage classification module is used to perform step S6 in claim 1; The second-stage subdivision and integration module is used to perform step S7 in claim 1; An interpretable display and interaction module is provided for performing step S8 of claim 1.

9. The system according to claim 8, characterized in that, It also includes a deployment and interface module, which is used to deploy the system to the hospital endoscopy workstation or server in the form of software service or hardware integration, and to interface with the hospital information system for data exchange.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the diagnostic method for vocal cord leukoplakia as described in any one of claims 1-7.

Images