A Cervical Cell Pathological Slide Classification Method Based on Weakly Supervised Learning

By constructing a multi-granularity information closed-loop cervical cell pathology slide classification system, and using feature entropy and activation heat to screen key image blocks, combined with attention weight calibration and hard pseudo-label refinement, the system solves the problems of insufficient instance-level localization accuracy and data imbalance in cervical cell pathology slide classification, and achieves efficient fine-grained lesion localization and classification.

CN122090165APending Publication Date: 2026-05-26HEFEI UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
HEFEI UNIV OF TECH
Filing Date
2026-03-20
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

Existing weakly supervised learning methods for classifying cervical cell pathology slides suffer from insufficient instance-level localization and classification accuracy, inadequate two-way knowledge flow, data class imbalance, and strong labeling dependence, making it difficult to meet the needs of fine-grained lesion localization.

Method used

A cervical cell pathology slide classification system based on weakly supervised learning is adopted. Through image preprocessing module, teacher branch module, student branch module and bidirectional optimization module, a multi-granularity information closed loop is constructed. Key image blocks are screened by feature entropy and activation heat. Combined with attention weight calibration, hard pseudo-label refinement and difficult positive instance mining, instance-level supervision and knowledge distillation are achieved.

Benefits of technology

It improves fine-grained classification and localization capabilities, alleviates the problem of data class imbalance, reduces reliance on manual annotation, enhances the ability to identify difficult positive instances, and improves the model's recognition accuracy and efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122090165A_ABST
    Figure CN122090165A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of image classification technology, specifically involving a cervical cell pathology slide classification method based on weakly supervised learning. First, key image patches are screened using a dual index of feature entropy and activation heat, and mapped to instance feature sets. The teacher branch calculates instance attention weights based on package-level labels, outputs package-level predictions after weighted aggregation, and generates soft pseudo-labels by normalizing the weights. The student branch fits the distribution of soft pseudo-labels through knowledge distillation and generates hard pseudo-labels. After fusing distillation and cross-entropy loss, the shared encoder parameters are updated. Finally, the updated encoder parameters are synchronized to the feature extraction stage. The model is iteratively optimized through alternating training by the teacher and student branches and a difficult instance mining mechanism, outputting classification results and generating a heatmap of positive instance location. This invention achieves high-precision instance-level classification and positive region location under weak supervision using only slide-level labels, effectively improving the identification ability of difficult positive instances.
Need to check novelty before this filing date? Find Prior Art