A pathological image instance segmentation method based on partial point labeling contrast learning

By constructing an interactive architecture and a dual-branch deep neural network model, and combining Markov decision processes and point-aware discrete contrastive learning, the high annotation cost and confirmation bias of existing cell nucleus segmentation methods during sparse-to-dense expansion are solved, achieving efficient and accurate pathological image segmentation.

CN122336286APending Publication Date: 2026-07-03BEIJING JIAOTONG UNIV
0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
BEIJING JIAOTONG UNIV
Filing Date
2026-04-03
Publication Date
2026-07-03

AI Technical Summary

Technical Problem

Existing cell nucleus segmentation methods suffer from high annotation costs, large confirmation biases when expanding from sparse to dense, and neglect of sequence decision-making and supervisory signal dilution, resulting in insufficient efficiency and accuracy in pathological image segmentation.

Method used

A contrastive learning method based on partial point annotation is adopted. By constructing an interactive architecture that includes a segmentation network and a policy network, a dense point label set is generated using a Markov decision process, and a dual-branch deep neural network model is constructed for training. Combined with point-aware discrete contrastive learning and block alignment mechanism, efficient pathological image segmentation is achieved.

Benefits of technology

Under sparse point annotation conditions, high-precision segmentation of cell nuclei was achieved, overcoming the confirmation bias caused by heuristic thresholding, solving the problem of supervisory signal dilution, and improving the accuracy and efficiency of pathological image segmentation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122336286A_ABST
    Figure CN122336286A_ABST
Patent Text Reader

Abstract

This invention provides a pathological image instance segmentation method based on partial point label contrastive learning. The method includes: preprocessing the original pathological image to be segmented to obtain a preprocessed pathological image and an initial sparse point label set; constructing an interactive architecture including a segmentation network and a policy network, using the segmentation network to generate a target probability prediction heatmap, and using the policy network to obtain an enhanced dense point label set; extracting positive and negative samples from the dense point label set; constructing a dual-branch deep neural network model including a representation learning branch and a segmentation branch, training the dual-branch deep neural network model using positive and negative samples, and outputting the instance segmentation result of the original pathological image to be segmented from the trained dual-branch deep neural network model. This invention achieves accurate pathological image segmentation by intelligently generating dense point labels through a reinforced guidance mechanism and combining point-aware discrete contrastive learning to enhance the feature recognition of positive and negative samples.
Need to check novelty before this filing date? Find Prior Art