CN-122023888-A - Gleason automatic grading for hybrid multi-instance learning based on enhanced hierarchical attention mechanisms

CN122023888ACN 122023888 ACN122023888 ACN 122023888ACN-122023888-A

Abstract

The invention provides Gleason automatic grading based on mixed multi-example learning of an enhanced hierarchical attention mechanism, and relates to the technical field of artificial intelligence and medical image analysis. The system comprises a multi-example pathology data preprocessing module, an enhanced hierarchical attention feature extraction module, a mixed multi-example learning classification module, a Gleason score mapping and optimizing module and a model iteration updating module. The preprocessing module divides a prostate cancer pathological slice into image block sets to construct a slice-image block two-level multi-example data structure, the hierarchical attention enhancement module associates attention with packet level time sequence through block level space-channel cooperation attention to extract focus key characteristics, and the mixed multi-example learning module fuses an MI-SVM and an attention weighting strategy to realize classification. The grading accuracy rate of the invention is more than or equal to 96%, the Kappa value consistent with the advanced pathologist is more than or equal to 0.88, the generalization is improved by 20%, the quality improvement and efficiency improvement of the pathological diagnosis can be assisted, and the invention is applicable to the clinical prostate cancer pathological diagnosis scene.

Inventors

REN MEILI
LIAN FENG
GUO QINGLI
REN MEIYAN
LI TING
TIAN MIAO

Assignees

山西财经大学

Dates

Publication Date: 20260512
Application Date: 20260120

Claims (8)

1. The Gleason automatic grading system for the mixed multi-example learning based on the enhanced hierarchical attention mechanism is characterized by comprising a multi-example pathology data preprocessing module, an enhanced hierarchical attention feature extraction module, a mixed multi-example learning classification module, a Gleason score mapping and optimizing module and a model iteration updating module, wherein the modules are sequentially in communication connection through a data interface; the multi-example pathology data preprocessing module constructs slice-image block two-level multi-example data; the enhanced hierarchical attention feature extraction module extracts hierarchical focus features; the hybrid multi-example learning classification module implements Gleason classification; the Gleason score mapping and optimizing module generates an optimized Gleason score; And the model iterative updating module realizes dynamic improvement of the model performance.
2. The Gleason auto-rating system for hybrid multi-example learning based on an enhanced hierarchical attention mechanism of claim 1, wherein the multi-example pathology data preprocessing module comprises: The slice segmentation unit is used for segmenting the pathological section into 512 multiplied by 512 pixel image blocks by adopting a sliding window method, and the overlapping rate is 20% -30%; the data normalization unit is used for carrying out gray level normalization and size normalization on the image block; And the abnormality screening unit is used for eliminating invalid image blocks with the background ratio more than or equal to 80% by an Otsu threshold method and constructing a plurality of example data sets with slices as 'packets' and valid image blocks as 'examples'.
3. The Gleason auto-rating system for hybrid multi-example learning based on an enhanced hierarchical attention mechanism of claim 1, wherein the enhanced hierarchical attention feature extraction module comprises a block-level enhanced attention unit and a package-level associated attention unit; a block-level attention-enhancing unit, which adopts cavity convolution to extract space features, combines the improved SE channel attention to generate space-channel collaborative attention force diagram, and weights the image block features; And the packet-level associated attention unit is used for sequencing the image block features of the same slice according to the spatial positions, inputting the bidirectional LSTM extraction time sequence associated features, generating packet-level attention weights through a self-attention mechanism and realizing the hierarchical aggregation of the block-level features.
4. The Gleason auto-classification system for hybrid multi-instance learning based on an enhanced hierarchical attention mechanism of claim 1, wherein the hybrid multi-instance learning classification module comprises: a basic classification unit, which takes an improved MI-SVM as a basic model, takes package-level aggregation characteristics as input, and outputs a preliminary classification result; the attention weighting unit fuses the block level attention weight and the packet level attention weight into a confidence factor and carries out weighted correction on the preliminary grading result; The loss optimizing unit adopts a joint function of cross entropy loss and center loss, the cross entropy loss optimizes classification boundaries, the center loss reduces the similar characteristic distance, and the weight ratio of the loss function is 7:3.
5. The Gleason auto-rating system of hybrid multi-instance learning based on an enhanced hierarchical attention mechanism of claim 1, wherein the Gleason score mapping and optimization module comprises: a rule mapping unit for mapping the grading result to an initial grade based on WHO2022 release Gleason grading standard; The Bayesian optimization unit takes the doctor scores of 1000 labeling samples as references, builds a score correction model, and adjusts the initial scores through maximum posterior probability estimation so that the score error is less than or equal to 0.3 score; and the result output unit is used for outputting the grading result, the corresponding score and the confidence coefficient.
6. The Gleason automatic grading system based on the mixed multi-example learning of the enhanced hierarchical attention mechanism according to claim 2, wherein the preprocessing module further comprises a data expansion unit, wherein an image block for generating focus morphological variation by using an countermeasure generation network GAN is generated, the expansion ratio is 1:2, a generator of GAN adopts a U-Net structure, a discriminator adopts a PatchGAN structure, and the structural similarity of the generated image block and a real sample is not less than 0.85.
7. The Gleason auto-rating system for hybrid multi-instance learning based on an enhanced hierarchical attention mechanism of claim 1, wherein the model iterative update module comprises: the incremental learning unit adopts a knowledge distillation strategy to transfer the knowledge of the historical model to a new model, so as to avoid catastrophic forgetting; and the on-line updating triggering unit automatically starts the fine adjustment of the model when the grading error of the newly accessed sample is more than or equal to 0.5 minute or the new pathological subtype appears, and the fine adjustment learning rate is 1/10 of the initial learning rate.
8. The Gleason automatic grading system based on mixed multi-example learning of the enhanced hierarchical attention mechanism according to any one of claims 1 to 7, wherein the system is adapted to HE-stained and immunohistochemical-stained prostate cancer pathological sections, digital pathological section WSI and local image block input are supported, the time for grading single WSI is less than or equal to 30s, the grading accuracy on a multi-center dataset is more than or equal to 96%, the recall is more than or equal to 95%, and the consistency Kappa value with more than 3 advanced pathologists is more than or equal to 0.88.

Description

Gleason automatic grading for hybrid multi-instance learning based on enhanced hierarchical attention mechanisms Technical Field The invention relates to the technical field of artificial intelligence and medical image analysis, in particular to Gleason automatic classification based on hybrid multi-example learning of an enhanced hierarchical attention mechanism. Background The Gleason grading and grading are "gold standards" formulated by prostate cancer diagnosis, prognosis evaluation and treatment schemes, and the core is that the malignancy degree of cancer cells is judged through the glandular structure morphology in pathological sections, the grade is divided into 3-5 grades, the corresponding grading is 3-5 grades, and finally the sum of the grades of the main grading and the secondary grading is taken as the total grading. However, the existing Gleason classification and scoring mode has significant limitations: 1. The subjective dependence is strong, the consistency is poor, the traditional dependence is greatly influenced by experience, fatigue and cognitive difference by naked eyes of a pathologist, the grading coincidence rate of different doctors on the same slice is only 65-75%, and the consistency Kappa value of primary hospitals and trimethyl hospitals is as low as 0.6, so that misdiagnosis or overstreated treatment is easily caused. 2. The multi-example data association mining is insufficient, namely pathological section WSI pixels reach billions, and the pathological section WSI pixels need to be divided into image blocks (examples) for analysis, so that a slice-image block multi-example data structure is formed. In the existing method, the image blocks are independently processed, and the spatial position association (such as focus infiltration trend) and semantic association (such as gland shape gradual change) of the image blocks in the same slice are ignored, so that key focus features are lost. 3. The attention mechanism has single level and rough feature extraction, the existing attention-based method only focuses on local features (such as single gland) in an image block, lacks hierarchical design of block level-packet level, cannot distinguish the weight of a key focus block from a normal tissue block, cannot aggregate global features of slice level, and has a hierarchical error rate of more than 10%. 4. The suitability of multi-example learning and grading tasks is poor, namely a traditional MI-SVM and other multi-example learning models can only output 'package' grading results, cannot be mapped to a Gleason multi-grade grading system accurately, and the classification boundary is optimized without combining the attention weight, so that the accuracy rate is suddenly reduced by more than 30% when the model is generalized to a new pathological subtype. In the prior art, as disclosed in patent (CN 202310567890.2), the 'deep learning-based Gleason classification method' only adopts single-stage attention to extract image block characteristics and does not construct hierarchical association, and patent (CN 202211456781.3) adopts a traditional MI-SVM to realize classification and lacks an optimization mechanism of score mapping. Therefore, an automatic grading and scoring scheme capable of mining multi-example data hierarchy association and accurately extracting focus features is needed, and core pain points with subjective dependence and insufficient accuracy are solved. Disclosure of Invention Aiming at the defects of the prior art, the invention provides the automatic Gleason grading based on the mixed multi-example learning of the enhanced hierarchical attention mechanism, and solves the problems of poor grading consistency, poor generalization and low precision caused by subjective judgment of pathologists, insufficient mining of block-packet hierarchical association and spatial semantic association in multi-example pathology data, single attention mechanism level, incapability of accurately positioning key focus features, and poor suitability of a multi-example learning and Gleason grading system. In order to achieve the above purpose, the invention is realized by the following technical scheme: The Gleason automatic grading system based on the mixed multi-example learning of the enhanced hierarchical attention mechanism comprises a multi-example pathology data preprocessing module, an enhanced hierarchical attention feature extraction module, a mixed multi-example learning classification module, a Gleason score mapping and optimizing module and a model iteration updating module, wherein the modules are sequentially in communication connection through a data interface; the multi-example pathology data preprocessing module constructs slice-image block two-level multi-example data; the enhanced hierarchical attention feature extraction module extracts hierarchical focus features; the hybrid multi-example learning classification module implements Gleason classification; the Gleaso