CN-122000076-A - Endometrial cancer molecular typing prediction system and method for cross-slice type generalization

CN122000076ACN 122000076 ACN122000076 ACN 122000076ACN-122000076-A

Abstract

The invention belongs to the technical field of medical artificial intelligence and digital pathology, and particularly relates to a cross-slice type generalization endometrial cancer molecular typing prediction system and method. The method comprises an image data acquisition and preprocessing module, an image block feature extraction and multi-scale fusion module, a model training module, a domain perception ViT aggregator and a domain perception attention mechanism, wherein the image data acquisition and preprocessing module acquires an endometrium cancer HE staining digital full-slice pathological image comprising preoperative, intraoperative and postoperative multi-slice types, the image block feature extraction and multi-scale fusion module acquires an image block feature sequence by carrying out multi-scale feature fusion on the image blocks, the model training module inputs the image block feature sequence into the domain perception ViT aggregator and utilizes the domain perception attention mechanism to adaptively adjust the attention of different slice types, and the domain perception feature aggregation and classification module predicts the slice-level molecular typing. The invention can overcome the interdomain difference of different pathological section types and realize rapid and accurate prediction of endometrial cancer molecular typing in preoperative and intraoperative stages.

Inventors

LIU QINGQING
Jiang Xuji
HAO YIPING
CUI BAOXIA

Assignees

山东大学齐鲁医院

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (10)

1. A cross-slice type generalization endometrial cancer molecular typing prediction system comprising: The image data acquisition and preprocessing module is configured to acquire digital full-slice pathological images of endometrial cancer HE staining, including preoperative, intra-operative and postoperative multi-slice types, and preprocess the digital full-slice pathological images to obtain a plurality of image blocks; the image block feature extraction and multi-scale fusion module is configured to fuse the multi-scale features of a plurality of image blocks to obtain an image block feature sequence; the model training module is configured to input the image block feature sequence into a domain perception ViT aggregator, adaptively adjust the attention of different slice types by using a domain perception attention mechanism, and complete the training of the HistoEMC-ViT model; The domain-aware feature aggregation and classification module is configured to input the endometrium cancer HE staining digital whole-section pathological images of the preoperative and/or intraoperative section types to be predicted into a trained HistoEMC-ViT model, and predict and obtain section-level molecular typing.
2. The cross-slice type generalized endometrial cancer molecular typing prediction system of claim 1, wherein in said image data acquisition and preprocessing module: The HE staining of the endometrial cancer before operation digitizes a full-section pathological image, in particular to a biopsy slice before operation; HE staining of endometrial cancer in operation digitizes a full-section pathological image, specifically an intraoperative frozen section; the postoperative endometrial cancer HE staining digital full-section pathological image is specifically a postoperative paraffin section.
3. The cross-slice type generalized endometrial cancer molecular typing prediction system of claim 1, wherein the image data acquisition and preprocessing module comprises the following specific preprocessing steps: Based on morphological operation and connected domain analysis, virtually associating and splicing the physically adjacent tissue fragments in a feature space, providing more complete tissue context information for a semantic segmentation model, and identifying an effective tissue region to obtain an initial image block; Introducing a quality evaluation network, grading the quality of the initial image blocks obtained by segmentation, automatically filtering the image blocks with the score lower than a preset threshold value, and reserving the image blocks with the score higher than the preset threshold value, wherein the preset threshold value is comprehensively set based on the extrusion degree or the ice crystal coverage rate.
4. The cross-slice type generalized endometrial cancer molecular typing prediction system of claim 3, wherein in the image block feature extraction and multi-scale fusion module, the multi-scale feature fusion is performed on a plurality of image blocks, and specifically comprises: Inputting the image block subjected to adaptive filtering into a pre-training Vision Transformer; Extracting features from the image blocks by Vision Transformer, extracting feature images from the intermediate layers, carrying out global average pooling, and extracting final classification token features to obtain features with different depths; and splicing the features with different depths to obtain an image block feature sequence.
5. The cross-slice type generalization endometrial cancer molecular typing prediction system of claim 2, wherein in the model training module, the generalization ability of the displayed optimization model by the improved meta-learning training strategy specifically comprises: Domain migration simulation, namely dividing a postoperative paraffin section data set into a plurality of meta-tasks in the training process, randomly selecting one meta-task in each iteration, further dividing the selected meta-task into a support set and a query set, simulating a known paraffin domain by using the support set, and simulating an unknown biopsy/freeze domain by using the query set; Double-loop optimization, namely, firstly performing one-step gradient update on a support set, namely, inner loop, and then calculating loss on a query set and performing final gradient update, namely, outer loop.
6. The cross-slice type generalized endometrial cancer molecular typing prediction system of claim 5, wherein in the model training module, a weighted cross entropy loss function is used to alleviate the problem of class imbalance in the training data: Loss = - Σ_{c=1}^C w_c Y_i,c log( _i,c); wherein C is the category number, Y_i, C is the one-hot encoding of the real label, I, c is the model predictive probability, and w_c is the weight for category c.
7. The cross-slice type generalized endometrial cancer molecular typing prediction system of claim 2, wherein said domain aware ViT aggregator comprises a domain aware attention mechanism and a standard fransformer layer, wherein: The domain aware attention mechanism is a learnable bias term related to slice type introduced in the self-attention mechanism of ViT aggregator, which is selectively superimposed on the attention score according to whether paraffin, ice or biopsy type is entered, so that the model can adaptively adjust the attention pattern; the standard transducer layer is stacked from multiple layers of transducer encoders, each layer of transducer encoder including multiple heads of self-attention, layer normalization, residual connection, and feed forward neural network.
8. The cross-slice type generalization endometrial cancer molecular typing prediction system of claim 1, further comprising a visualization module configured to: Developing a dual-channel visualization system, and displaying a self-attention-based heat map and an integral gradient-based attribution heat map in parallel; Wherein the self-attention based heat map shows the image block region that the model focuses on when making the decision, and the integral gradient based attribution heat map shows the pixel level region that contributes most to the final prediction.
9. The cross-slice type generalization endometrial cancer molecular typing prediction system of claim 5, wherein the training mode of the dual-loop optimization forces the model to learn feature representations and decision boundaries that are not only valid on a single paraffin slice data distribution, but remain robust against data distribution changes, i.e., domain shifts.
10. A method for predicting endometrial cancer molecular typing across-slice type generalization, comprising the steps of: Acquiring digital full-section pathological images of endometrial cancer HE staining including preoperative, intraoperative and postoperative multi-section types, and preprocessing to obtain a plurality of image blocks; Carrying out multi-scale feature fusion on a plurality of image blocks to obtain an image block feature sequence; Inputting the image block feature sequence into a domain perception ViT aggregator, and adaptively adjusting the attention of different slice types by using a domain perception attention mechanism to finish training a HistoEMC-ViT model; and (3) inputting the pre-operation and/or intra-operation slice type endometrial cancer HE staining digital full-section pathological images to be predicted into a trained HistoEMC-ViT model, and predicting to obtain the slice-level molecular typing.

Description

Endometrial cancer molecular typing prediction system and method for cross-slice type generalization Technical Field The invention belongs to the technical field of medical artificial intelligence and digital pathology, and particularly relates to a cross-slice type generalization endometrial cancer molecular typing prediction system and method. Background Molecular typing of endometrial cancer is a cornerstone for accurate diagnosis and treatment. However, the prior art has the following fundamental limitations: (1) The prediction time is seriously delayed, the molecular typing detection (such as NGS and IHC) of gold standard depends on a paraffin specimen after operation, which takes days to weeks, and the result is available after the operation is completed and cannot be used for guiding the formulation of a preoperative scheme (such as operation scope and lymph node cleaning necessity) and the decision in operation (such as operation type adjustment). (2) The domain limitation of the existing AI model is that the currently published molecular parting model based on deep learning is completely based on postoperative paraffin sections for training, verification and application. The performance of these models on biopsies and frozen sections is drastically reduced due to the large inter-domain differences between paraffin sections and preoperative biopsies (small amounts of tissue, fragmentation, with crush injuries), intra-operative frozen (ice crystal artifacts, cell morphology deformations) sections, and the inability to apply directly. (3) There is no specific generalization solution, and no research or technology has been disclosed to effectively solve the problem of domain generalization from paraffin sections to biopsies/frozen sections. The method is directly applied to general methods such as field adaptation or data enhancement, and has limited effect on the specific and complex medical image domain offset problem. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides a cross-slice type generalization endometrial cancer molecular typing prediction system and a cross-slice type generalization endometrial cancer molecular typing prediction method, which can overcome the domain difference among different pathological slice types and realize rapid and accurate prediction of endometrial cancer molecular typing in preoperative and intraoperative stages. To achieve the above object, one or more embodiments of the present invention provide the following technical solutions: The first aspect of the invention provides a cross-slice type generalization endometrial cancer molecular typing prediction system. A cross-slice type generalization endometrial cancer molecular typing prediction system comprising: The image data acquisition and preprocessing module is configured to acquire digital full-slice pathological images of endometrial cancer HE staining, including preoperative, intra-operative and postoperative multi-slice types, and preprocess the digital full-slice pathological images to obtain a plurality of image blocks; the image block feature extraction and multi-scale fusion module is configured to fuse the multi-scale features of a plurality of image blocks to obtain an image block feature sequence; the model training module is configured to input the image block feature sequence into a domain perception ViT aggregator, adaptively adjust the attention of different slice types by using a domain perception attention mechanism, and complete the training of the HistoEMC-ViT model; The domain-aware feature aggregation and classification module is configured to input the endometrium cancer HE staining digital whole-section pathological images of the preoperative and/or intraoperative section types to be predicted into a trained HistoEMC-ViT model, and predict and obtain section-level molecular typing. In a second aspect, the invention provides a method for prediction of endometrial cancer molecular typing across slice type generalizations. A method for prediction of endometrial cancer molecular typing across-slice type generalization, comprising the steps of: Acquiring digital full-section pathological images of endometrial cancer HE staining including preoperative, intraoperative and postoperative multi-section types, and preprocessing to obtain a plurality of image blocks; Carrying out multi-scale feature fusion on a plurality of image blocks to obtain an image block feature sequence; Inputting the image block feature sequence into a domain perception ViT aggregator, and adaptively adjusting the attention of different slice types by using a domain perception attention mechanism to finish training a HistoEMC-ViT model; and (3) inputting the pre-operation and/or intra-operation slice type endometrial cancer HE staining digital full-section pathological images to be predicted into a trained HistoEMC-ViT model, and predicting to obtain the slice-level molecular typi