Cervical lesion thermodynamic diagram prediction method and system based on pathological base model fine tuning, storage medium and program

Abstract

The invention provides a cervical lesion thermodynamic diagram prediction method and system based on pathological base model fine tuning, a storage medium and a program, and relates to the technical field of medical pathological diagnos.The method comprises the steps that S10, a pre-trained Prov-GigaPath tissue pathological base model is adopted to extract pathological feature vectors of a cervical tissue panoramic digital slice image; s20, executing dichotomy mapping according to the pathological feature vectors, and generating positive probabilities of the sub-regions; s30, generating a highly suspicious sub-region set according to the positive probability of each sub-region; and S40, executing five-classification prediction according to the height suspicious sub-region set, and generating lesion level probability distribution corresponding to each height suspicious sub-region. Compared with the prior art, the method has the following beneficial effects: 1, two-classification and five-classification tasks are decoupled, so that the model complexity is reduced; secondly, fine-grained distinguishing of CINI-III and canceration is achieved through a five-classification grading framework; thirdly, by embedding an uncertainty quantification mechanism, a lesion grading result and model confidence degree evaluation are synchronously output.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical pathological diagnosis, in particular to a cervical lesion heat map prediction method and system based on pathological base model fine-tuning, a storage medium and a program. BACKGROUND

[0002] In the field of medical image diagnosis, the existing technology mainly relies on traditional artificial microscope reading or single-stage classification framework based on deep learning. Traditional artificial reading has the problems of strong subjectivity, long time consumption and being easily affected by the experience difference of pathologists, especially for the grading of complex lesions such as cervical intraepithelial neoplasia (CIN) (such as CIN1 to CIN3 and cancer), which has poor diagnostic consistency and high risk of missed detection. In recent years, single-stage classification methods based on deep learning directly classify whole slide images (WSI) through full connection layers or convolutional neural networks (CNN), which realizes preliminary screening of lesions by extracting image features and training classifiers, and to some extent, assists clinical decision-making.

[0003] Although the above method improves the automation level, however, most models still focus on the two-classification framework (normal / lesion) or the simplified three-classification framework (low-grade / high-grade / cancer), and fail to design a fine-grained five-classification logic for the continuous pathological changes of CIN I-III and cancer. Such simplified framework breaks the pathological progressive relationship between lesion grades, resulting in the inability of the output heat map to accurately distinguish between mild dysplasia (such as CIN I) and significant dysplasia (such as CIN III), thereby failing to provide effective basis for clinical stratified intervention such as regular follow-up of CIN I and immediate treatment of CIN III.

[0004] In addition, the existing model does not introduce a quantitative mechanism for predicting uncertainty when outputting lesion grade probabilities, i.e., the model's confidence assessment of the results. For example, when the model predicts a probability of CIN II of 70%, if no confidence interval or uncertainty indicator is provided for the result, the clinician cannot determine whether the prediction needs to be manually reviewed, which directly weakens the clinical practical value of the auxiliary diagnosis system. Although the existing heat map can mark suspicious areas, its probability distribution does not combine the dynamic correlation of multi-scale features, and lacks feedback on the reliability of the model's own prediction, resulting in the heat map being unable to distinguish between lesion areas of different severity, and also being unable to support the grading decision of clinicians. SUMMARY

[0005] In view of the above problems existing in the prior art, the technical problem to be solved by the present application is to provide a cervical lesion heat map prediction method and system based on pathological base model fine-tuning, a storage medium and a program, which can realize end-to-end judgment from cervical panoramic section to fine-grained lesion grading and confidence assessment.

[0006] To solve the above technical problems, the technical solution adopted by the present invention is as follows: the present invention provides a cervical lesion heat map prediction method based on fine-tuning of the pathological base model, comprising the following steps: S10. Use the pre-trained Prov-GigaPath histopathology base model to extract the pathological feature vector of the panoramic digital slice image of cervical tissue ; S20, according to the pathological feature vector Perform binary classification mapping to generate positive probabilities for subregions ; S30, based on the positive probability of each sub-region Generate a set of highly suspicious sub-areas ; S40, based on the highly suspicious sub-area collection Perform five-category prediction to generate the probability distribution of lesion grade corresponding to each highly suspicious sub-region ; S50, obtaining uncertainty threshold and uncertainty indicators , compared with the uncertainty index and uncertainty threshold Determine the lesion grading results.

[0007] In a preferred solution, a composite function is obtained according to steps S10, S20, S30, S40 and S50: ; in, To input the panoramic digital slice image of cervical tissue, is the pathological feature extraction function in step S10, is the positive probability of the sub-region in step S20 The calculation function of is the threshold screening formula, is the function for the five-category prediction and uncertainty calculation in steps S40 and S50, is the function composition operator.

[0008] In the preferred solution, a binary classification task labeling system is constructed: Negative label = 0, indicating chronic cervicitis; positive label = 1, indicating CIN-Ⅰ, CIN-Ⅱ, CIN-Ⅲ and cancer; Determine the five-category task label system based on the positive class label: Chronic cervicitis is mapped to label 0; CIN-I is mapped to label 1; CIN-II is mapped to label 2; CIN-III is mapped to label 3; cancer is mapped to label 4; Determine the five-category label k according to the five-category task label system and map the five-category label k to the output neuron .

[0009] In the preferred solution, the specific steps of step S20 are: S21, pathological feature vector Input to the linear classification head to perform binary classification mapping; Among them, the input dimension of the linear classification head is set to d, and the output dimension is set to 2, corresponding to two neurons respectively and ; Map negative class labels to output neurons , Represents the lesion; maps the positive class label to the output neuron , represents non-lesional; S22, using the binary cross entropy loss function to optimize the classification parameters; S23. Update the weight matrix of the linear classification head through SGD iteration and bias , minimize the cross entropy loss ; S24. Pathological feature vector for each sub-region Perform forward inference to calculate the positive probability of each sub-region , output the positive probability matrix of all sub-regions in the panoramic digital slice image of cervical tissue .

[0010] In a preferred solution, the cross entropy loss function formula in step S22 is: ; Where n is the total number of sub-regions, i is the sub-region index, is the true label of the i-th sub-region, i.e., 0 = non-lesion, 1 = lesion; For neurons The corresponding probability is the probability of lesion; For neurons The corresponding probability is the probability of non-lesion.

[0011] In the preferred solution, in step S30, a dynamic threshold of positive probability is preset. and ,filter The sub-regions are regarded as highly suspicious sub-regions, and a set of highly suspicious sub-regions is generated. ; in, is the probability of positive test Output a value in the range of 1; In step S40, the pathological feature vector In step S40, the pathological feature vector The input five-classifier performs a five-classification transformation, and the output dimension is 5, wherein a Dropout layer is embedded in the five-classifier, the Dropout is enabled in the inference stage, and a cross-entropy loss function is used to optimize the parameters of the five-classifier SGD is used to iteratively update the parameters The output is a trained "five-classifier with Dropout layer"; Each pathological feature vector In step S40, the pathological feature vector is input into the five-classifier with Dropout layer to perform independent forward reasoning times, and each time a single-sampling five-classification probability distribution is obtained ; The average five-classification probability distribution is determined according to all single-sampling five-classification probability distributions , and the formula is: ; Wherein, is the average probability of the i-th region belonging to the lesion grade at the t-th sampling.

[0012] In the preferred scheme, the specific steps of step S50 are as follows: S51, calculate the probability fluctuation of the i-th region belonging to the label k in T times of prediction, and the variance formula of the single-label probability is: ; Wherein, is the probability of the i-th highly suspicious sub-region belonging to the k-th class at the t-th sampling; The variance vector formula of each lesion grade probability is: ; Take the maximum value in it as the uncertainty index, and the formula is: ; Output the binary tuple for each highly suspicious sub-region; Set the uncertainty threshold ; When , it is marked as "high confidence", and is directly adopted as the lesion grading result; When , it is marked as "recommend manual review", and , and the sub-region coordinates are output, and a manual review task is triggered through the hospital PACS system.

[0013] In a preferred implementation, the present application also provides a cervical lesion thermogram prediction system based on pathological base model fine-tuning, characterized in that the system for performing any of the above-mentioned cervical lesion thermogram prediction methods based on pathological base model fine-tuning comprises: an image input and preprocessing module for receiving cervical tissue panoramic digital slice images and performing size adjustment, center cropping and standardization processing to generate a set of standardized sub-region tiles , wherein is output to the feature extraction module; a feature extraction module for receiving the set of sub-region tiles output by the image input and preprocessing module , extracting a set of high-dimensional pathological feature vectors through a pre-trained Prov-GigaPath model , wherein is output to the binary classification prediction module; a binary classification prediction module for receiving the set of high-dimensional pathological feature vectors output by the feature extraction module , performing binary classification prediction through a linear classifier to generate a full-image positive probability matrix , wherein is output to the suspicious area screening module; a suspicious area screening module for receiving the positive probability matrix output by the binary classification prediction module , screening out a set of highly suspicious sub-regions based on a dynamic threshold , wherein is output to the five-class prediction module; a five-class prediction module for receiving the set of highly suspicious sub-regions output by the suspicious area screening module , calculating a five-class probability distribution through multiple rounds of forward sampling and Monte Carlo Dropout technology , wherein is output to the uncertainty quantification module; an uncertainty quantification module for receiving the five-class probability distribution output by the five-class prediction module , calculating an uncertainty index , wherein and are output to the result integration module; a result integration and decision module for receiving and output by the uncertainty quantification module, combining a preset threshold to generate an "automatic adoption / manual review" decision, and outputting a full-image analysis report to the visualization and system interface module; The visualization and system docking module is used for receiving the full picture analysis report, and outputting the heat map, hierarchical result and review mark in the full picture analysis report to the visualization terminal.

[0014] In a preferred solution, the application further provides a computer non-transitory readable storage medium, which stores computer programs / instructions, characterized in that the computer programs / instructions are executed by a processor to implement the steps of the cervical lesion heat map prediction method based on the pathological base model fine-tuning according to any one of the above.

[0015] In a preferred solution, the application further provides a computer program product, which comprises computer programs / instructions, characterized in that the computer programs / instructions are executed by one or more processors to implement the steps of the cervical lesion heat map prediction method based on the pathological base model fine-tuning according to any one of the above.

[0016] The application provides a cervical lesion heat map prediction method based on a pathological base model fine-tuning, and compared with the prior art, the method, system, storage medium and program have the following beneficial effects: First, the binary classification and five-classification tasks are decoupled to reduce the model complexity, and the feature vector reuse is used to reduce data redundancy, wherein the dynamic threshold adjustment based on the positive probability can reduce redundant calculation through a screening mechanism, optimize resource allocation, and be suitable for real-time diagnosis requirements in a clinical scene. Second, the five-classification hierarchical architecture is used to realize fine-grained differentiation of CIN I-III and cancer, solve the problem of lack of multi-level lesion evaluation in a traditional binary classification model, make the heat map accurately reflect the severity of the lesion, and support clinical stratified intervention decisions through the generated fine-grained heat map, and significantly improve the pathological level resolution and clinical applicability of diagnosis. Third, the uncertainty quantification mechanism is embedded to synchronously output the lesion grading result and model confidence evaluation, and according to the confidence threshold, it is determined whether manual review is needed, reduces high-risk missed diagnosis (i.e., low-confidence result misjudgment) and excessive medical treatment (i.e., high-confidence result repeated review), and improves the reliability of the auxiliary decision system. BRIEF DESCRIPTION OF DRAWINGS

[0017] The application will be further described below in combination with the drawings and embodiments: Figure 1 is a main view structural diagram of the flow of the application; Figure 2 is a five-classification flow structural diagram of the application; Figure 3 is a system logic block diagram of the application; Figure 4 is a heat map of 30 typical cases in Embodiment 4 of the application; Figure 54 is a comparison chart of the prediction results of Case 4 in Example 4 of the present invention; Figure 6 This is a comparison chart of the prediction results of Case 5 in Example 4 of the present invention; Figure 7 This is a comparison chart of the prediction results of Case 6 in Example 4 of the present invention; Figure 8 It is a structural schematic diagram of the computer device of the present invention. DETAILED DESCRIPTION

[0018] In order to better understand the purpose, system architecture and functional implementation of this embodiment, the embodiments and features in the embodiments of this application can be combined with each other without conflict. The following will describe the exemplary embodiments disclosed in this embodiment in conjunction with the accompanying drawings, which contain specific technical details disclosed in this embodiment to assist understanding, but these details should be regarded as exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the present invention. Similarly, for the sake of clarity, detailed descriptions of well-known technologies, functions and structures (such as standard image processing algorithms, general communication protocols) are omitted in the following description.

[0019] In the field of medical pathology diagnosis technology, artificial intelligence-assisted diagnosis has gradually been applied to cervical lesion screening, but there are still significant deficiencies in fine-grained grading accuracy, diagnostic efficiency, model confidence quantification and clinical process closure.

[0020] With the development of digital pathology technology, the widespread use of panoramic cervical sections (WSIs) provides a data foundation for AI analysis. Pre-trained pathology base models can learn common features from large-scale pathology data, significantly reducing the annotation cost of specific tasks. However, the diagnosis of cervical lesions requires accurate differentiation of fine-grained levels such as chronic cervicitis, CIN-I / II / III, and cancer. Existing solutions have the following core limitations: Inadequate grading capabilities: Traditional AI models focus on the binary classification of "lesion / non-lesion" and are unable to distinguish between CIN-I / II / III grades, such as the extent of nuclear atypia and the depth of epithelial invasion. This results in a lack of accurate grading basis for treatment plans. Prominent efficiency bottlenecks: Directly performing fine-grained analysis on the entire image with millions of pixels requires enormous computational effort, meaning that processing a single case takes over 5 minutes, making it difficult to adapt to the large-scale screening scenarios of "hundreds of cases per day" in primary hospitals. Unquantified confidence: The model only outputs the probability of classification and does not quantify the "uncertainty of the prediction." This means that the probability of CIN-II and CIN-III fluctuates in the critical range, making it difficult for clinicians to determine whether manual review is necessary. This increases the risk of missed diagnosis (e.g., missing CIN-III) or misdiagnosis (e.g., misdiagnosing CIN-I as CIN-II). Therefore, the prior art has not yet constructed an integrated solution of "fine-grained classification → efficiency optimization → confidence quantification → clinical closed loop", which is difficult to meet the stringent requirements of pathological diagnosis on precision, efficiency, explainability and process coordination.

[0021] Embodiment 1 In this embodiment, a prov-gigapath histopathological base model based on pathological base model fine-tuning for cervical lesion thermogram prediction method is constructed, and the specific training and parameter setting are as follows: S1, multi-dimensional pathological data acquisition and labeling system construction Obtain cervical tissue images determined by multiple rounds of pathologists. This dataset covers five typical pathological states of cervical tissue, including chronic cervicitis (normal / low risk), CIN-I (mild squamous intraepithelial lesion), CIN-II (moderate squamous intraepithelial lesion), CIN-III (severe squamous intraepithelial lesion), and cancer (invasive carcinoma), ensuring the clinical representativeness of lesion classification. All images come from clinical biopsy samples, processed into high-resolution digital images by Whole Slide Imaging (WSI) technology, and then cut into ROI (region of interest) tiles containing typical lesions according to pathological diagnosis standards, finally forming 400,000 labeled training images.

[0022] Among them, the stratified sampling strategy commonly used in clinical data is adopted, and the training set (360,000 images) and the verification and test set (40,000 images) are divided according to the 9:1 ratio, ensuring that each lesion class is evenly distributed in different datasets. Each image is independently labeled by more than 3 pathologists, and expert review procedures are initiated for labeling disagreement samples to construct a high-precision labeling benchmark (labeling consistency ≥98%) to provide reliable supervision signals for subsequent model training S2, data preprocessing Size uniformity: Resize the original pathological tiles to 224x224 pixels to match the input requirements of mainstream visual models and eliminate size differences caused by different scanning devices; Center cropping: To address potential interference from edge area artifacts, a center area priority cropping strategy is adopted to preserve the core lesion area; Color standardization: Z-score standardization (mean / variance normalization) is used to eliminate the impact of staining differences (such as HE staining depth) on model feature extraction, and data augmentation techniques such as random flipping, rotation, brightness / contrast perturbation are simultaneously applied to improve model generalization ability; Tissue mask filtering: Use semantic segmentation technology to remove irrelevant blank background and non-epithelial tissue areas in the image, focusing on the effective pathological tissue area.

[0023] Label conversion: Convert the diagnosis results of doctors into digital labels that the model can recognize. The specific mapping relationship is as follows: (1) Binary classification task label system: Negative class (label 0): Chronic cervicitis (low risk, representing no / low-grade lesions); Positive class (label 1): CIN-I, CIN-II, CIN-III, and cancer (high-risk lesions, representing different degrees of high-grade lesions and malignant transformation).

[0024] This label system not only retains the core distinction of clinical grading (normal vs. lesions), but also provides an extension interface for subsequent fine-grained five-class upgrade.

[0025] (2) Based on the positive class (label 1), the five-class task label system is as follows: Chronic cervicitis is mapped to label 0; CIN-I is mapped to label 1; CIN-II is mapped to label 2; CIN-III is mapped to label 3; and cancer is mapped to label 4.

[0026] S3, AI algorithm selection: Use prov-gigapath histopathology base model; Prov-GigaPath is a histopathology base model pre-trained based on real data from 28 US hospitals of Providence, covering 30,000 patients. The training data includes 170,000 whole slide digital pathology images (WSI) and 1.3 billion pathology tiles, covering a wide range of pathological scenarios.

[0027] Specifically, the advantages of using the prov-gigapath histopathology base model lie in its excellent performance in 26 pathological tasks, including 9 cancer classification tasks and 17 pathological tasks; In 25 tasks, it achieved leading results, significantly outperforming existing methods in 18 tasks, demonstrating its strong ability to extract histological features, including deep learning of histomorphology, cell atypia, and stromal reaction. The pre-training parameter size reaches hundreds of millions, and a general feature representation space for the pathology field is constructed.

[0028] This model captures invariant features between different staining methods (HE, immunohistochemistry) and imaging devices through a contrast learning mechanism, and has strong sensitivity to key lesion indicators such as cervical tissue nuclear polarity, mitotic figures, and epithelial layer disorder; In the 26-pathological task benchmark test, 25 tasks reached the SOTA (state-of-the-art) level, and 18 tasks significantly outperformed contemporary methods. In particular, in the cervical precancerous lesion identification task, the feature extraction accuracy improved by 19.7% compared to traditional CNN, providing a strong feature expression foundation for downstream diagnostic tasks.

[0029] S4, heat map prediction model training, through a two-stage training framework, as follows: Feature extraction layer: freeze the first 12 layers of Prov-GigaPath pre-training parameters (preserve general tissue features), only fine-tune the last 6 layers of channel weights, adapt to the specific lesion patterns of cervical pathological images (such as the basal cell hyperplasia characteristics of squamous intraepithelial lesions); Coarse classification decision layer: access LinearClassifier, build a binary classification model, output pixel-level positive probability, set its range to [0, 1], and the specific design is as follows: Label mapping: chronic cervicitis is mapped to 0, CIN-I / II / III+ cancer is mapped to 1, focusing on the preliminary screening of "whether there is a high-risk lesion"; Loss function: Cross-EntropyLoss, to strengthen the decision boundary of high and low risk lesions: ; Optimization strategy: SGD optimizer (momentum = 0.9), batch_size = 2048 (balance video memory and gradient stability); Learning rate lr = 0.02 (cosine annealing decay), weight decay = 0.01 (to suppress overfitting); Train for 8000 iterations (about 100 epochs), achieve 95.2% training accuracy and 93.8% validation accuracy on 400,000 samples, and the generated heat map can locate the core area of ≥90% CIN-II and above lesions.

[0030] Specifically, the panoramic heat map is generated by Grad-CAM technology, and the high positive probability area is marked, providing a "suspicious area ROI set" for the following step S5 model, solving the low efficiency problem of traditional model "blind full image analysis".

[0031] S5, suspicious area disease prediction model training, i.e. fine-grained five-class lesion grading model training, and based on the heat map output by S4, extract high-suspicious area features with positive probability ≥ threshold, build an end-to-end architecture of "basin model feature extraction + five-class linear head" as follows: Input preprocessing: secondary standardization (mean / variance re-calibration) of ROI patches filtered from the heat map, to eliminate the interference of different area brightness differences.

[0032] Feature reuse and optimization: reuse Prov-GigaPath full network features (unfreeze all parameters), perform task-specific fine-tuning on the cervical lesion dataset, and strengthen the recognition ability of subtle differences in CIN-I to cancer (such as the full-layer cell nucleus atypia of CIN-III and the interstitial infiltration signal of cancer).

[0033] Five-class decision layer: label definition: 0 = chronic cervicitis, 1 = CIN-I, 2 = CIN-II, 3 = CIN-III, 4 = cancer, strictly corresponding to the clinical pathological grading standard; Loss function: multi-class cross-entropy loss, formula: wherein, is the true label (0 or 1) of the i-th sample in class c, is the predicted probability.

[0034] Training parameters and optimization: optimizer: continue the SGD algorithm (batch_size = 2048, lr = 0.02, weight_decay = 0.01), add EarlyStopping mechanism (stop if the validation set loss does not decrease for 50 consecutive iterations), avoid overfitting; Training strategy: use "warm-up" pre-training (linearly increase lr to 0.02 for the first 1000 iterations), gradually activate pre-trained parameters, improve the classification sensitivity of small sample classes (such as cancer, accounting for about 3%); Performance indicators: on the validation test set, the five-class accuracy is 89.7%, among which the CIN-II and CIN-III classification accuracy is 85.3%, which is significantly better than the direct five-class baseline model (78.2%), achieving the upgrade of fine-grained diagnosis from "whether there is a lesion" to "severity".

[0035] The five-class results output by the model can directly correspond to the clinical diagnosis and treatment guidelines (such as CIN-II and above requiring clinical intervention), combined with the suspicious area located by the S4 heat map, forming a complete diagnosis chain of "lesion positioning-grade judgment", providing pathologists with three-dimensional information of "region coordinates-lesion grading-probability confidence", solving the core pain points of traditional two-class models that cannot guide grading decisions.

[0036] The pre-training ability of the Prov-GigaPath model combined with the lightweight design of the linear classification head ensures that the model has high accuracy while being computationally efficient, suitable for large-scale pathological image analysis, combined with a phased classification strategy, significantly improving the accuracy and efficiency of cervical lesion detection.

[0037] Example 2 This example is based on the Prov-GigaPath base model fine-tuned on 400,000 cervical pathological images in Example 1, and constructs a four-level prediction process of "feature extraction-two classification-coarse screening positioning-fine-grained grading-uncertainty quantification".

[0038] As Figure 1 , 2 ​As shown, the system input is the cervical tissue whole section digital image obtained by clinical biopsy, and the specific processing steps of the cervical lesion heat map prediction method based on pathological base model fine-tuning are as follows: S10, prediction of the heat map prediction model: the pre-trained Prov-GigaPath histopathological base model is used to extract the pathological feature vector of the cervical tissue panoramic digital section image . The specific steps are as follows: S11, input the cervical tissue panoramic digital section image (full field) into the pre-trained Prov-GigaPath histopathological base model, and segment multiple sub-regions of the cervical tissue panoramic digital section image.

[0039] S12, according to the pre-trained Prov-GigaPath histopathological base model, the features of each sub-region in the input cervical tissue panoramic digital section image are extracted, the hierarchical pathological features are extracted through a deep convolutional neural network, and a high-dimensional pathological feature vector set is generated : , wherein ; wherein is the d-dimensional pathological feature vector of the i-th sub-region, i is the sub-region index (i=1, 2, …, n), n is the total number of sub-regions after segmentation of the panoramic image, d is the feature dimension, is a d-dimensional real number space; S20, according to the pathological feature vector perform binary classification mapping to generate the positive probability of the sub-region ; The specific steps are as follows: S21, input the pathological feature vector to the linear classification head (linear head classifier) to perform binary classification mapping; wherein the input dimension of the linear classification head is set to d, which is consistent with the output dimension of the feature extraction module, and the output dimension is set to 2, corresponding to two neurons and , that is represents a lesion, represents a non-lesion; Chronic cervicitis is set as label 0, and label 0 is mapped to output neuron ; CIN-I / II / III / cancer is set as label 1, and label 1 is mapped to output neuron .

[0040] S22, the classification parameters are optimized using a binary cross-entropy loss function (Cross-Entropy Loss), and the cross-entropy loss function formula is: ; where n is the total number of sub-regions, i is the sub-region index, is the true label of the i-th sub-region, i.e. 0 = non-disease, 1 = disease; is the probability corresponding to the neuron , i.e. the disease probability; is the probability corresponding to the neuron , i.e. the non-disease probability.

[0041] Specifically, the core of the direct mapping binary classification defined by the label is to distinguish "disease (label 1)" from "non-disease (label 0)", where "disease" includes CIN-I / II / III and cancer, i.e. positive regions that need to be focused on clinically. Therefore, the neuron directly corresponds to the "disease" category, and its probability naturally represents the positive probability .

[0042] S23, the weight matrix and the bias of the linear classification head are updated iteratively by stochastic gradient descent (SGD) or its variants (i.e. Adam) to minimize the cross-entropy loss ; S24, the pathological feature vector of each sub-region is subjected to forward inference to calculate the positive probability of each sub-region , and the steps are as follows: S241, the linear classification head outputs the original logit: ; where is the original output (logit) of the classification head, is the weight matrix, is the pathological feature vector, is the bias vector.

[0043] S242, after Softmax normalization, the probability distribution is obtained: ; where Softmax is the normalization function.

[0044] S243, the positive probability is defined as the probability corresponding to the disease category (i.e. label 1) as the positive probability , i.e. ; Specifically, since Softmax satisfies , therefore Directly reflects the confidence of "lesion", ranging from [0,1].

[0045] Finally, the positive probability matrix of all sub-regions in the panoramic digital slice image of cervical tissue is output ,The matrix elements correspond one-to-one to the global coordinates of the sub-regions in the WSI, providing a spatial mapping for subsequent suspicious area screening.

[0046] S30. Prediction of suspicious area disease prediction model: The positive probability of each sub-area of ​​the panoramic digital slice image of cervical tissue is obtained based on the heat map prediction model Generate a set of highly suspicious sub-areas ; Based on highly suspicious areas Perform classification prediction and output the probability distribution of lesion grade corresponding to each highly suspicious sub-region ; The specific steps are: S31. Preset dynamic threshold of positive probability and , preferably, the positive probability threshold The threshold data is calibrated by clinical data, balancing the missed diagnosis rate and computational efficiency, and screening out The sub-regions of are regarded as highly suspicious sub-regions, forming a highly suspicious sub-region set , at the same time, only Execute five categories to reduce 80% to 85% of invalid calculations and improve system efficiency.

[0047] in, is the probability of positive test Outputs values ​​in the range 1.

[0048] S32, will Pathological feature vector in Input five classifiers to perform five-class transformation; Among them, the output dimension of the five classifiers is 5, and the corresponding labels are set as follows: Chronic cervicitis is set as label 0, which means it is not a lesion and needs to be excluded; CIN-Ⅰ was set as label 1, i.e., low-grade lesion; CIN-Ⅱ was set as label 2, indicating intermediate-grade lesions; CIN-III is set as label 3, which means high-grade lesions and precancerous lesions; Canceration is set as label 4, i.e., malignant lesions; Map label 0 to the output neuron , mapping label 1 to the output neuron , mapping label 2 to the output neuron , mapping label 3 to the output neuron mapping the label 4 to the output neuron .

[0049] In summary, the mapping of the neurons is as follows: let k be a five-class label, and k correspond to the output neuron of the five-class classifier , realizing a linear mapping from “feature to hierarchy”.

[0050] S33, embedding a Dropout layer in the five-class classifier, keeping the Dropout enabled in the training or inference stage (Monte Carlo Dropout mechanism), so that the model learns “robust features under random inactivation”, avoiding overfitting and enhancing the generalization ability; Specifically, through multiple rounds of backpropagation, the suspicious area disease prediction model parameters are adapted to the disturbance caused by Dropout, enhancing the ability to depict uncertainty, and combining the SGD optimizer to update the parameters.

[0051] But because of the existence of Dropout, the gradient of each backpropagation comes from “random subnetwork”, and finally the model parameters will be more “anti-disturbance”, providing a basis for multiple sampling in the inference stage.

[0052] Using cross-entropy loss function to optimize the parameters of the five-class classifier , the formula of the cross-entropy loss function is: ; Wherein, is the true label of sub-region i for label k, and one-hot encoding is performed, such as CIN-II corresponding to , the rest are 0, is the predicted probability of the five-class classifier in the training stage that i belongs to k, without sampling, directly forward output.

[0053] Then, using stochastic gradient descent (SGD) or Adam, iteratively update the parameters .

[0054] Output the trained “five-class classifier with Dropout layer”.

[0055] S34, input the pathological feature vector in step S31 and the trained “five-class classifier with Dropout layer” in step S33; for each pathological feature vector in , keep the Dropout layer enabled, and repeat the execution of Second, the dropout layer randomly deactivates different neurons during each inference, simulating the prediction of multiple sub-models, so that a five-class probability distribution is obtained for each sampling. The formula for the five-class probability distribution of a single sampling is: ; where t = 1, 2,..., T represents the t-th sampling, and the five neurons correspond to five lesion grades, i.e. = chronic cervicitis, = CIN-I, = CIN-II, = CIN-III, = cancer. By aggregating the T sampling results, the random fluctuations are reduced, and the normalized probability distribution is output by the Softmax function. The formula for the average prediction probability of the model that the region belongs to the lesion grade is: ; where is the probability that the i-th highly suspicious sub-region belongs to the k-th class at the t-th sampling, i.e., the prediction result under the random deactivation of the dropout.

[0056] The average prediction probability of the lesion grade is ; where is the probability that the i-th region belongs to the lesion grade at the t-th sampling.

[0057] Specifically, by aggregating the prediction results of the T sub-models, the "consensus judgment" of the model for the region is obtained, reducing the random fluctuations caused by a single dropout, making more stable and reliable, and serving as the core basis for lesion grading.

[0058] In this embodiment, in addition to the lesion grade probability that needs to be predicted, the uncertainty index also needs to be calculated. When the uncertainty index exceeds the threshold, an artificial review task is automatically triggered, the diagnosis cycle is shortened, and the clinical value of the technical solution is strengthened. In simple terms, it is to quantify the model confidence: First, calculate the variance of each lesion grade probability, i.e., the formula for the variance of the single-label probability is: ; where is the probability that the i-th region belongs to label k at the t-th sampling, i.e., the prediction result under the random deactivation of the dropout.

[0059] Construct the variance vector, the formula is: ; Take the maximum value as the uncertainty index, the formula is: ; Wherein, is the probability fluctuation degree of the i-th region belonging to label k in T times of prediction, that is, the larger the variance, the more uncertain the model, indicates the fluctuation degree of the model to the most uncertain level among all lesion levels of the i-th region.

[0060] Specifically, if is large, the maximum value is taken, for example, if the prediction results of a certain region frequently fluctuate between CIN-II and CIN-III, it indicates that the model is inconsistent in grading judgment for this region, and the uncertainty is high; On the contrary, if is small, it indicates that the model is stable in judgment and has high confidence.

[0061] It should be noted that the traditional AI only outputs "determined grading", and the present scheme quantifies the "uncertainty degree" through Monte Carlo Dropout, for example It prompts the CIN-II / CIN-III criticality, directly drives manual review, and avoids the risk of "algorithm confidence but actual misjudgment".

[0062] S34, output a binary tuple ; Further, according to clinical verification, a preset uncertainty threshold is preferred, and the uncertainty threshold is 0.3, which is calibrated by 400,000 difficult cases to ensure that more than 95% of critical lesions are captured.

[0063] And calibrated by the expert-labeled difficult case set as follows: When , it is marked as "model high confidence", and is directly adopted as the final lesion grading result and automatically written into the pathology report; When , it is marked as "suggested manual review", and , and sub-region coordinates are output, and the hospital PACS system triggers the manual review task, and the pathologist further confirms, for example, the CIN-II and CIN-III critical region needs to be confirmed by the pathologist.

[0064] The above steps S10, S20 and S30 are combined to obtain a composite function: ; in, To input the panoramic digital slice image of cervical tissue (raw data), The pathological feature extraction function in step S10 is used to output the pathological feature vector , is the positive probability of the sub-region in step S20 The calculation function of Used to generate a set of highly suspicious sub-areas , is the function of five classifications and uncertainty calculation, is the five-category parameter of step S30, T is the probability distribution of aggregated T samplings, is the function composition operator.

[0065] The final lesion grading results are visualized by overlaying the five-category grading colors on a heat map and connected to the hospital PACS system through the DICOM / DICOMweb interface; Among them, if the final lesion grading result is cancerous, it will be marked in red, and if the final lesion grading result is CIN-III, it will be marked in orange.

[0066] Example 3 The following describes a cervical lesion heat map prediction system based on fine-tuning of a pathological base model provided by the present invention. The cervical lesion heat map prediction system based on fine-tuning of a pathological base model described below and the cervical lesion heat map prediction method based on fine-tuning of a pathological base model described above can be referred to each other and further explained in conjunction with Examples 1 and 2. like Figure 3 The structure shown, Figure 3 A cervical lesion heat map prediction system based on fine-tuning of a pathological base model is provided in an embodiment of the present application. This embodiment consists of seven core functional modules, and each module realizes data flow-driven collaborative work through a standardized data interface, including an image input and preprocessing module, a feature extraction module, a two-category prediction module, a suspicious area screening module, a five-category prediction module, a result integration and uncertainty quantification module, and a visualization output module. The modules are electrically connected to each other and work together through data flow and control flow to realize an end-to-end process from whole slice image (WSI) input to lesion heat map generation, graded prediction, and clinical decision support.

[0067] In this embodiment, module 1 is an image input and preprocessing module, which is used for image preprocessing, converting clinical full-slice pathology images into standardized blocks that can be processed by the model, and completing physical segmentation and noise filtering.

[0068] Specifically, image parsing: support input of full-slice image (WSI) in DICOM / SVS / TIFF format, parse image metadata (resolution, magnification, etc.); Chunk processing: using sliding window technology, preferably, window size 224x224 pixels, overlap rate 30%, no missing segmentation, WSI is divided into n continuous sub-region image blocks, generating a sub-region tile set ; Color normalization: calibrate pixel values by Z-score algorithm, mean 0.45, variance 0.25, eliminate staining differences; Invalid filter: use tissue mask to remove blocks with background ratio > 70%, record global coordinates (x, y) for each sub-region, establish a mapping relationship between "sub-region-global position", output image block set with coordinates, and output effective block set; Data formatting: group effective blocks by preset batch size, preferably, batch size is 1024 / batch, generate input tensor (dimension: Nx224x224x3) for model inference.

[0069] In this embodiment, module 2 is a feature extraction module, which is used to output original feature vectors, and based on a pre-trained Prov-GigaPath base model, high-dimensional pathological feature vectors of sub-regions are extracted.

[0070] Specifically, deep feature extraction: the sub-region tile set is input into the fine-tuned Prov-GigaPath model, forward propagation is performed, and hierarchical pathological features (such as cell nucleus morphology and epithelial structure) are extracted through a convolutional neural network to generate a d-dimensional feature vector, d is a 2048-dimensional deep feature vector, which retains multi-level pathological features such as cell nucleus morphology and epithelial structure; Feature caching: store the feature vector according to the block coordinate index, and output the feature set to the binary classification prediction module and the suspicious region screening module.

[0071] In this embodiment, module 3 is a binary classification prediction module, which is used to perform "lesion / non-lesion" coarse classification on the sub-regions of the whole image, generate a positive probability matrix, and screen a candidate set of high-suspicious regions.

[0072] Specifically, binary classification head inference: the feature vector is mapped to the positive probability through a linear classification layer (i.e. parameters ), and the positive probability is calculated through a Softmax function, and a whole image positive probability matrix is constructed according to the sub-region coordinates.

[0073] In this embodiment, module 4 is a suspicious region screening module, which is used to screen high-suspicious sub-regions according to the positive probability, extract corresponding feature vectors, and narrow the scope of subsequent analysis.

[0074] Specifically, spatial aggregation: based on the coordinates of suspicious patches, merge adjacent suspicious regions by connected component analysis (4 / 8 neighborhood connection) to generate an initial set of highly suspicious sub-regions.

[0075] In this embodiment, module 5 is a five-class prediction module, which is used to process the highly suspicious sub-regions output by the suspicious region screening module, and to perform fine-grained classification on the suspicious and highly suspicious sub-regions to achieve accurate classification of CIN I-III and cancer.

[0076] Specifically, dynamic threshold screening: set a positive probability threshold to screen sub-regions, and form a set of highly suspicious sub-regions ; Feature subset extraction: from the feature set of module 2, extract the pathological feature vector corresponding to the sub-region , and record its global coordinates.

[0077] Input the positive probability matrix of the whole image (from module 3) and the sub-region feature vector (from module 2); Output the set of highly suspicious sub-region feature vectors and the corresponding coordinates, and pass them to the five-class prediction and uncertainty quantification module.

[0078] In this embodiment, module 5 is a five-class prediction and uncertainty quantification module, which is used to perform fine-grained classification (CIN-I~III and cancer) on highly suspicious sub-regions and to quantify the prediction uncertainty through Monte Carlo Dropout.

[0079] Specifically, a linear classification head with embedded Dropout layer is used, where the output dimension is 5, Dropout is enabled during training (Monte Carlo Dropout mechanism), and Dropout is kept enabled during inference; Multiple rounds of forward sampling: for each , perform independent forward inference, each time with different neurons randomly inactivated by Dropout, to obtain T five-class probability distributions ( ); Probability aggregation: calculate the average probability distribution of T samplings as the classification result of the region, where labels 0-4 correspond to chronic cervicitis, CIN-I~III, and cancer; Uncertainty calculation: calculate the variance of the probability of each lesion grade: ; Take the maximum variance as the uncertainty indicator: , reflecting the confidence of model classification. With the input of high-suspected sub-region feature vectors (from module 4); , the classification probability of each sub-region , the uncertainty indicator and the corresponding coordinates are output to the result integration and clinical decision module.

[0080] In this embodiment, module 6 is the result integration and decision module, which is used to aggregate the classification results of high-suspected regions, combine the uncertainty indicators, generate clinical diagnosis decisions, and automatically adopt or manually review the suggestions.

[0081] Specifically, if , it is marked as "high confidence" and the highest probability level of is directly adopted; if : marked as "manual review suggestion", the coordinates of the highlighted region, probability distribution and ; Global integration: aggregate all decision results of high-suspected sub-regions to generate a full-image diagnostic report, including classification, confidence, and review suggestion.

[0082] Input the , and coordinates of each sub-region (from module 5), and the preset threshold ; Output the full-image diagnostic report, i.e. structured data, to the visualization and system integration module.

[0083] In this embodiment, module 7 is the visualization and system integration module, which is used to visualize the diagnostic results (including heat map and region annotation) and integrate with the hospital information system to trigger the clinical process.

[0084] Specifically, according to the positive probability matrix , a full-image heat map (red-coded high-suspected regions) is constructed, and five-class classification colors are superimposed, such as red for cancer and orange for CIN-III; Report visualization: superimpose the , and decision suggestions of each high-suspected region on the heat map in the form of annotations to generate an interactive view; System integration: upload the diagnostic report and heat map to the hospital PACS system through the DICOM / DICOMweb interface; if there is a "manual review" mark, trigger the task flow of the hospital information system to notify the pathologist for review.

[0085] Input the full-image positive probability matrix (From module 3), full picture diagnostic report (from module 6); Output clinical diagnostic report, including visual heat map, interface to hospital information system, complete diagnostic closed loop.

[0086] Example 4 This embodiment is based on the cervical lesion heat map prediction method based on pathological base model fine-tuning of example 2, 30 typical samples are selected from 400,000 labeled complete training image data, that is, 30 cases are extracted from LD0000001 to LD0402389. All samples are confirmed by pathological experts through strict manual film reading to ensure the authority and accuracy of sample positive and negative judgment.

[0087] In the experiment, the model of example 1, the algorithm of example 2 and the system of example 3 are used for automatic analysis of the above samples. At the same time, the analysis results of this scheme are compared with the prediction results of the single inference model of the traditional method to comprehensively evaluate the effectiveness and reliability of this scheme in medical image quality evaluation.

[0088] The following will combine specific experimental data to elaborate the specific implementation process of this scheme in medical image quality evaluation in detail, and show the corresponding experimental results, which provide data support and practical basis for the practical application of this scheme.

[0089] As shown in the following table 1, 30 cases of cervical tissue whole section images (WSI) diagnosed by pathological gold standard are selected, covering CIN-I (5 cases), CIN-II (10 cases), CIN-III (10 cases), and cancer (5 cases), to verify the system whole process running logic.

[0090] Table 1 Diagnosis results of high suspicious sub-region

[0091] As Figure 4 shown, the heat map of 30 data in table 1 is shown, which shows the positive probability distribution of different cases. In this embodiment, the critical lesion cases marked as "yes" by artificial review are preferentially analyzed, that is, , which directly reflects the core ability of the algorithm in "fine-grained classification" and "uncertainty quantification".

[0092] Case 1: LD1567890, CIN-III critical, artificial review triggered; Original data features: positive probability , that is, it is determined as a high suspicious lesion, and the five classification result is 3 (CIN-III), but the probability distribution is scattered (CIN-II: 45%, CIN-III: 48%), and the uncertainty .

[0093] Algorithm output vs. human verification: Algorithm marked “suggested manual review”, pathologist reviewed and confirmed as CIN-III, with local nuclear hyperchromasia but no clear atypia.

[0094] Analysis conclusion: Through uncertainty quantification, critical lesions are correctly identified, avoiding misdiagnosis due to close probabilities. Traditional binary classification models may directly classify as “lesions” without distinguishing grades. Case 2: LD2890123, CIN-II evolving into a cancerous critical lesion, triggering manual review. Raw data features: Positive probability , five-class result is 2 (CIN-II), but cancer probability is 1%, uncertainty , as CIN-II and CIN-III probabilities are close: 79% vs. 8%, variance amplifies uncertainty.

[0095] Manual verification conclusion: Review found individual cell polarity loss at lesion edge, ultimately revised to CIN-III, algorithm missed edge lesions due to focusing on central area features.

[0096] Analysis conclusion: Monte Carlo Dropout captures subtle fluctuations in probability distribution, triggering review even if five-class result is clear, reducing missed diagnosis risk. Case 3: LD3012345, CIN-III, manual review triggered but actually cancerous. Raw data features: Positive probability , five-class result is 3 (CIN-III), cancer probability is only 4%, but uncertainty , i.e., five-class probability distribution is uniform: 84% CIN-III, 7% CIN-II, 4% cancer.

[0097] Manual verification conclusion: Review found lesion with stromal invasion, ultimately diagnosed as early-stage cancer, algorithm missed malignant transformation due to lack of stromal feature learning.

[0098] Analysis conclusion: Uncertainty indicator effectively identifies model “unconfidence” cases, avoiding missed low-probability cancer as CIN-III.

[0099] In this embodiment, we prioritize cases marked “no” for manual review, i.e. , accurately predicted lesion cases, from Table 1, we select 2 cases predicted as cancer (five-class = 4) and 1 case predicted as CIN-II (five-class = 2), and conduct in-depth analysis combining data features, algorithm output, and clinical logic.

[0100] For example, Figure 5Case 4: LD0134365, typical cancer, high confidence recognition.

[0101] Raw data features: Positive probability: 0.98, close to the theoretical maximum, the model is highly suspicious of the lesion; Five-category distribution: cancer probability 92%, the rest <3%, probability distribution is highly concentrated; Uncertainty: Monte Carlo sampling variance is minimal, and the model is highly confident in the prediction.

[0102] Algorithm output and manual verification: the algorithm directly outputs "cancer, confidence 92%" and does not trigger manual review; Pathologist verification: there are a large number of mitotic figures, loss of cell polarity, and interstitial infiltration in the lesion, and it is diagnosed as invasive carcinoma.

[0103] Advantages: deep feature sensitivity: Prov-GigaPath extracts cancer-specific features such as nuclear polymorphism and nuclear-cytoplasmic ratio >0.7, which have formed strong discriminative ability in pre-training; Probability distribution is concentrated: the five-category head strengthens the boundary between cancer and other categories through cross-entropy loss, avoiding low probability misjudgment.

[0104] As Figure 6 shown in Case 5: LD2234567, a borderline cancer, multiple features are used for comprehensive judgment; Raw data features: positive probability: 0.96 (highly suspicious lesion); Five-category distribution: cancer probability 90%, CIN-III probability 4%, slightly higher than other benign categories; Uncertainty: Monte Carlo sampling variance is minimal, and no review is triggered.

[0105] Algorithm determines "cancer, confidence 90%" and is directly adopted; Pathological verification: there are local CIN-III features at the edge of the lesion, but the core area has broken through the basement membrane, and it is diagnosed as early cancer.

[0106] Advantages: The model combines nuclear deep staining (CIN-III feature) and interstitial micro-infiltration (cancer feature), and preferentially adopts cancer-specific signals (such as basement membrane destruction) through weight distribution. Monte Carlo Dropout reduces model overconfidence, so even if there are a few interference features, the probability distribution is still stable through multiple sampling. Potential defects: If the cancer feature is not typical (such as micro-infiltration range <1mm), the probability may fluctuate slightly due to the low proportion of early cancer samples in the training data (5%), but in this case, no review is triggered.

[0107] As Figure 7 shown in case 6: LD0345678, a typical CIN II, fine-grained classification verification; Original data characteristics: positive probability: 0.82 (definitely suspicious lesion); Five-category distribution: CIN II probability 85%, CIN-I and CIN-III probability 4% and 7% respectively, distribution concentrated; Uncertainty: (The model is more confident about the classification result).

[0108] The algorithm output is "CIN II, confidence 85%", without manual review; Pathological verification: abnormal cells can be seen in 1 / 3~2 / 3 of the epithelial layer, and mitotic figures are located in the upper layer, which meets the CIN II diagnostic criteria.

[0109] Advantages: The model captures the core feature of CIN II - abnormal cells occupying 1 / 2~2 / 3 of the epithelial layer, which is different from CIN I (1 / 3) and CIN III (full layer), and realizes accurate classification through spatial feature positioning; Clear category boundary: the five-category architecture forces the distinction between the five levels, avoiding the ambiguity of the traditional two-category "lesion" label, and directly corresponding to the clinical treatment guidelines (such as CIN II requiring conization or close follow-up); Defects: If the proportion of abnormal cells in the epithelial layer is close to the CIN-III critical value (such as 2 / 3~3 / 4), the probability may fluctuate due to feature overlap, such as case 15's CIN II / CIN-III probability is close, but in this case, the feature is typical and there is no misjudgment.

[0110] In this embodiment, as shown in the recognition accuracy comparison table of Table 2, the scheme solves the core defect of the traditional method "only can judge whether there is a lesion, can not be subdivided level", through five-category architecture and uncertainty quantification, cancer detection specificity is improved significantly, avoid CIN-III misjudged as cancer, traditional method is easy to overestimate the risk because of strong cell atypia; The traditional method is completely missing in the CIN-II / III distinction, while the present scheme achieves a distinction rate of 85.3% through fine-grained feature learning, but still has 14.7% of the critical cases that need manual review, see cases 1, 2.

[0111] Table 2 Recognition accuracy comparison table

[0112] In this embodiment, as shown in the processing efficiency comparison table of Table 3, the scheme compresses the single instance processing time to 1 / 3 of the traditional method through the "rough screening positioning to fine-grained classification" architecture, and the lesion positioning is completely automated, eliminating the time-consuming of manual preliminary screening; The manual review rate is reduced from 30% to 10% in the traditional method, but there is still room for optimization, such as case 3 triggered by insufficient interstitial characteristics, which can be further reduced by enhancing cancer sample training.

[0113] Table 3 Processing efficiency comparison table

[0114] In this embodiment, as shown in the following Table 4 abnormal cell detection ability table, the cancer detection accuracy (96%) and confidence (0.94) are the highest, which reflects the strong sensitivity of the model to malignant characteristics; The CIN-III missed diagnosis rate (13%) is higher than that of other grades, mainly because the characteristics such as nuclear-cytoplasmic ratio critical value are subtle and different from CIN-II, and nuclear quantification analysis needs to be increased.

[0115] Table 4 Abnormal cell detection ability table

[0116] In this embodiment, as shown in the following Table 5 manual intervention rate and pathological gold standard verification table, the manual review rate and the gold standard coincidence rate both reach the preset target, which proves that the scheme has achieved a balance between "reducing manual burden" and "ensuring diagnostic reliability"; The critical lesion coverage rate (95%) exceeds the preset threshold, thanks to the comprehensive capture of low confidence results by the uncertainty quantification mechanism, such as cases 1-3 being correctly labeled.

[0117] Table 5 Manual intervention rate and pathological gold standard verification table

[0118] In summary, the system realizes accurate grading diagnosis of cervical intraepithelial neoplasia (CIN-I~III) and cancer through an innovative five-classification framework, breaking through the limitations of the traditional two-classification "cancer / non-cancer" mode.

[0119] In terms of core performance, the system has three technical advantages: First, the fine-grained grading model based on deep learning achieves a cancer detection specificity of 98.2%, effectively avoiding the risk of over-treatment; Second, the innovative application of Monte Carlo Dropout uncertainty quantification mechanism reduces the manual review rate to 10% through the U_j index, while still covering 92.4% of potential misdiagnosis cases; Finally, the use of pipeline parallel architecture compresses the single instance processing time to 2.5 minutes while maintaining a five-classification accuracy of 89.7%, meeting the large-scale screening needs of primary medical institutions.

[0120] For continuous optimization of practical application scenarios, the system plans three improvement directions: by introducing interstitial infiltration, angiogenesis and other cancer-specific pathological features, the misdiagnosis rate of CIN-III and cancer can be reduced by 12%; With the help of knowledge distillation and MobileNet variants, the model is lightweight, which is expected to improve the inference speed by 40%; At the same time, a dynamic threshold calibration mechanism is established, combined with the diagnostic standard difference of each medical institution, The trigger threshold is intelligently adjusted from 0.35 in primary hospitals to 0.25 in third-grade hospitals, achieving the optimal balance between diagnostic efficiency and resource investment. The hierarchical technical architecture of the above embodiments not only guarantees the reliability of diagnosis, but also significantly improves the screening efficiency of primary medical care through intelligent decision support system.

[0121] Embodiment 5 Further illustrated in combination with Embodiment 1, as shown in the structure, Figure 8 The structure of the computer device provided by the embodiments of the present application is shown. The computer device includes: Figure 8 a processor, a memory, a communication bus, and a computer program stored in the memory and executable on the processor.

[0122] The processor can call the computer program in the memory, and implement the cervical lesion heat map prediction method based on pathological base model fine-tuning provided in the above embodiments when executing the program. The method includes: S10, using a pre-trained Prov-GigaPath histopathological base model to extract a pathological feature vector of a cervical tissue panoramic digital slice image ; S20, performing binary classification mapping according to the pathological feature vector to generate a positive probability of the sub-region ; S30, generating a highly suspicious sub-region set according to the positive probability of each sub-region ; S40, performing five-class prediction according to the highly suspicious sub-region set to generate a lesion grade probability distribution corresponding to each highly suspicious sub-region ; S50, obtaining an uncertainty threshold and an uncertainty index , comparing the uncertainty index and the uncertainty threshold to determine the lesion grading result.

[0123] Further, the computer device further includes: a communication interface (Communications Interface) for communication between the memory and the processor.​​

[0124] The memory can include a high-speed RAM memory and can also include a non-volatile memory, such as at least one disk memory.

[0125] If the memory, the processor and the communication interface are implemented independently, the communication interface, the memory and the processor can be connected to each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For convenience of representation, Figure 8 In the figure, only one thick line is used to represent, but it does not mean that there is only one bus or one type of bus.

[0126] In addition, the logical instructions in the above-mentioned memory can be implemented in the form of a software functional unit and sold or used as an independent product, and can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application or the part of the technical solutions which essentially contribute to the prior art or the part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0127] The processor can include one or more processing units, for example: the processor can include an application processor (AP), an application specific integrated circuit (ASIC), a modem processor, a central processing unit (CPU), an image signal processor (ISP), a controller, a memory, a video code, a digital signal processor (DSP), a baseband processor, and / or a neural-network processing unit (NPU), etc. Different processing units can be independent devices or integrated in one or more processors. The controller can be a nerve center and command center. The controller can generate operation control signals according to instruction operation codes and timing signals, and complete the control of fetching and executing instructions. A memory can also be provided in the processor to store instructions and data. In some embodiments, the memory in the processor is a cache memory. The memory can save instructions or data that the processor has just used or repeatedly uses. If the processor needs to use the instructions or data again, it can directly call them from the memory. This avoids repeated access and reduces the waiting time of the processor, thereby improving the efficiency of the system.

[0128] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0129] The display device is used for displaying images, videos, etc. The display device can include a display panel, which can adopt a liquid crystal display (LCD), an organic light-emitting diode (OLED), an active-matrix organic light-emitting diode (AMOLED), a flex light-emitting diode (FLED), a Miniled, a MicroLed, a Micro-oLed, a quantum dot light emitting diode (QLED), etc.

[0130] Optionally, in a specific implementation, if the memory, the processor and the communication interface are integrated on a chip, the memory, the processor and the communication interface can complete the communication among each other through an internal interface.

[0131] In another aspect, the embodiment of the present application also provides a computer non-transient readable storage medium, which stores a computer program, and the program is executed by a processor to implement the cervical lesion thermogram prediction method based on pathological base model fine-tuning as above, and the method comprises the following steps: S10, extracting a pathological feature vector of a cervical tissue panoramic digital slice image by using a pre-trained Prov-GigaPath histopathological base model ; S20, performing binary classification mapping according to the pathological feature vector to generate a positive probability of a sub-region ; S30, generating a highly suspicious sub-region set according to the positive probability of each sub-region ; S40, performing five-classification prediction according to the highly suspicious sub-region set to generate a lesion grade probability distribution corresponding to each highly suspicious sub-region ; S50, obtaining an uncertainty threshold value and an uncertainty index ; comparing the uncertainty index and the uncertainty threshold value to determine a lesion grading result.

[0132] ​On the other hand, the embodiment of the present application further provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium, and the computer program can run computer instructions. When the computer program is executed by a processor, the computer can execute the cervical lesion heat map prediction method based on fine-tuning of the pathological base model provided by the above methods, the method comprising: S10, using the pre-trained Prov-GigaPath tissue pathology base model to extract the pathological feature vector of the panoramic digital slice image of the cervical tissue ; S20, according to the pathological feature vector Perform binary classification mapping to generate positive probabilities for subregions ; S30, according to the positive probability of each sub-region Generate a set of highly suspicious sub-areas ; S40, based on the highly suspicious sub-area collection Perform five-category prediction to generate the probability distribution of lesion grade corresponding to each highly suspicious sub-region ; S50, obtain uncertainty threshold and uncertainty indicators , compared with the uncertainty index and uncertainty threshold Determine the lesion grading results.

[0133] The logic and / or steps represented in the flowchart or otherwise described herein may be considered, for example, as an ordered list of executable instructions for implementing logical functions, and may be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device).

[0134] For the purposes of this specification, a "computer-readable medium" can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-readable medium can be a computer- readable storage medium or a computer-readable communication medium. Computer-readable storage medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection having one or more wires (electronic apparatus), a portable computer diskette (magnetic apparatus), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber (optical apparatus), and a portable compact disc read-only memory (CDROM), optical storage device, magnetic storage device, or any suitable combination of the foregoing. Computer-readable storage medium can even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via, for instance, optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory.

[0135] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, specially designed application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0136] It should be understood that various forms of flow shown above can be used, with steps reordered, added, or removed. For example, steps recited in the present disclosure can be performed in parallel, in series, or in a different order, without limitation herein, so long as the desired results of the technology disclosed in the present disclosure are achieved.

Claims

1. A cervical lesion heat map prediction method based on fine-tuning of a pathological base model, characterized in that: The following steps are involved: S10. Use the pre-trained Prov-GigaPath histopathology base model to extract the pathological feature vector of the panoramic digital slice image of cervical tissue ; S20, according to the pathological feature vector Perform binary classification mapping to generate positive probabilities for subregions ; S30, based on the positive probability of each sub-region Generate a set of highly suspicious sub-areas ; S40, based on the highly suspicious sub-area collection Perform five-category prediction to generate the probability distribution of lesion grade corresponding to each highly suspicious sub-region ; S50, obtaining uncertainty threshold and uncertainty indicators , compared with the uncertainty index and uncertainty threshold Determine the lesion grading results.

2. The cervical lesion heat map prediction method based on fine-tuning of the pathological base model according to claim 1 is characterized in that: According to steps S10, S20, S30, S40 and S50, a composite function is obtained: ； in, To input the panoramic digital slice image of cervical tissue, is the pathological feature extraction function in step S10, is the positive probability of the sub-region in step S20 The calculation function of is the threshold screening formula, is the function for the five-category prediction and uncertainty calculation in steps S40 and S50, is the function composition operator.

3. The cervical lesion heat map prediction method based on fine-tuning of the pathological base model according to claim 1 is characterized in that: Build a binary classification task label system: Negative class label = 0, indicating chronic cervicitis; Positive class label = 1, indicating CIN-Ⅰ, CIN-Ⅱ, CIN-Ⅲ and cancer; Determine the five-category task label system based on the positive class label: Chronic cervicitis is mapped to label 0; CIN-I is mapped to label 1; CIN-II is mapped to label 2; CIN-III is mapped to label 3; cancer is mapped to label 4; Determine the five-category label k according to the five-category task label system and map the five-category label k to the output neuron .

4. The cervical lesion heat map prediction method based on fine-tuning of the pathological base model according to claim 3 is characterized in that: The specific steps of step S20 are: S21, pathological feature vector Input to the linear classification head to perform binary classification mapping; Among them, the input dimension of the linear classification head is set to d, and the output dimension is set to 2, corresponding to two neurons respectively and ; Map negative class labels to output neurons , Represents the lesion; maps the positive class label to the output neuron , represents non-lesional; S22, using the binary cross entropy loss function to optimize the classification parameters; S23. Update the weight matrix of the linear classification head through SGD iteration and bias , minimize the cross entropy loss ; S24. Pathological feature vector for each sub-region Perform forward inference to calculate the positive probability of each sub-region , output the positive probability matrix of all sub-regions in the panoramic digital slice image of cervical tissue .

5. The cervical lesion heat map prediction method based on fine-tuning of the pathological base model according to claim 4 is characterized in that: In step S22, the cross entropy loss function formula is: ； Where n is the total number of sub-regions, i is the sub-region index, is the true label of the i-th sub-region, i.e., 0 = non-lesion, 1 = lesion; For neurons The corresponding probability is the probability of lesion; For neurons The corresponding probability is the probability of non-lesion.

6. The cervical lesion heat map prediction method based on fine-tuning of the pathological base model according to claim 5 is characterized in that: In step S30, a dynamic threshold of positive probability is preset. and ,filter The sub-regions are regarded as highly suspicious sub-regions, and a set of highly suspicious sub-regions is generated. ; in, is the probability of positive test Output values ​​in the range of 1; In step S40, Pathological feature vector in The input five classifiers are transformed into five categories, and the output dimension is 5. The Dropout layer is embedded in the five classifiers. The Dropout layer is kept enabled during the inference phase, and the cross entropy loss function is used to optimize the parameters of the five classifiers. , use SGD to iteratively update the parameters , the output is the trained "five classifiers with Dropout layer"; Through the five classifiers with Dropout layer Each pathological feature vector in implement Independent forward inference, each sampling obtains the five-category probability distribution of a single sampling ; According to the five-category probability distribution of all single samples Determine the average five-category probability distribution , the formula is: ； in, The ith region belongs to the lesion level at the tth sampling The average probability of .

7. The cervical lesion heat map prediction method based on fine-tuning of the pathological base model according to claim 6 is characterized in that: The specific steps of step S50 are: S51. Calculate the probability fluctuation of the i-th region belonging to label k in T predictions. The variance formula of the single-label probability is: ； in, is the probability that the i-th highly suspicious sub-region belongs to the k-th category at the t-th sampling; The variance vector formula of each lesion grade probability is: ； The maximum value is taken as the uncertainty index, and the formula is: ； Output a binary group based on each highly suspicious sub-region ; Set uncertainty threshold ; when , marked as "high confidence", directly adopted As a result of lesion grading; when , marked as "manual review recommended", output 、 and sub-area coordinates, triggering a manual review task through the hospital PACS system.

8. A cervical lesion heat map prediction system based on fine-tuning of the pathological base model, characterized in that: A system for executing the cervical lesion heat map prediction method based on fine-tuning of a pathological base model according to any one of claims 1 to 7, comprising: Image input and preprocessing module, used to receive panoramic digital slice images of cervical tissue and perform size adjustment, center cropping and standardization to generate a standardized sub-region tile set ,Will Output to the feature extraction module; Feature extraction module, which receives image input and sub-region block set output by preprocessing module , extracting a high-dimensional pathology feature vector set through the pre-trained Prov-GigaPath model ,Will Output to the binary classification prediction module; The binary classification prediction module is used to receive the high-dimensional pathology feature vector set output by the feature extraction module. , perform binary classification prediction through linear classifier to generate the full image positive probability matrix ,Will Output to the suspicious area screening module; Suspicious area screening module, used to receive the positive probability matrix output by the binary classification prediction module , based on the dynamic threshold, filter out highly suspicious sub-region sets ,Will Output to the five-category prediction module; Five-category prediction module, used to receive the highly suspicious sub-region set output by the suspicious region screening module , calculate the five-category probability distribution through multiple rounds of forward sampling and Monte Carlo Dropout technology ,Will Output to the uncertainty quantification module; Uncertainty quantification module, used to receive the five-category probability distribution output by the five-category prediction module , calculate the uncertainty index ,Will and Output to the result integration module; The result integration and decision module is used to receive the output of the uncertainty quantification module. and , combined with the preset threshold, it generates "automatic adoption / manual review" decisions and outputs the full-map analysis report to the visualization and system docking module; The visualization and system docking module is used to receive the full-map analysis report and output the heat map, grading results and review marks in the full-map analysis report to the visualization terminal.

9. A computer non-transitory readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the method for predicting cervical lesion heat map based on fine-tuning of the pathological base model as described in any one of claims 1 to 7 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by one or more processors, the steps of the method for predicting cervical lesion heat maps based on fine-tuning of a pathological base model as described in any one of claims 1 to 7 are implemented.

Images