Related Experiment Video
Updated: May 1, 2026
10:59
Reconstruction of 3-Dimensional Histology Volume and its Application to Study Mouse Mammary Glands
Published on: July 26, 2014
14.3K
MSRMMP: Multi-scale residual module and multi-layer pseudo-supervision for weakly supervised segmentation of
Yuanchao Xue1, Yangsheng Hu1, Yu Yao1
1Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650504, China.
Medical Engineering & Physics
|February 20, 2025
Summary
This study introduces a new weakly supervised semantic segmentation (WSSS) method for histopathology images, significantly reducing annotation time and cost while improving cancer diagnosis accuracy.
Area of Science:
- Digital Pathology
- Computational Biology
- Medical Image Analysis
Background:
- Accurate semantic segmentation of histopathological images is vital for cancer diagnosis.
- Fully supervised methods achieve high performance but require extensive manual annotation.
- Weakly supervised semantic segmentation (WSSS) uses image-level labels to reduce annotation burden.
Purpose of the Study:
- To develop a novel two-stage weakly supervised segmentation framework (MSRMMP) for histopathological images.
- To address the limitations of existing WSSS methods in capturing multi-target information and pseudo-mask inaccuracy.
- To reduce the significant time and cost associated with manual image annotation.
Main Methods:
- Proposed a two-stage framework: MSRMMP, involving pseudo-mask generation and multi-layer pseudo-supervision.
- Utilized multi-scale residual networks (MSR-Net) with multi-scale residual modules (MSRM) for robust feature extraction and pseudo-mask generation.
- Employed TransUNet as the segmentation backbone and incorporated multi-layer pseudo-supervision to refine segmentation accuracy.
Main Results:
- The proposed MSRMMP method demonstrated superior performance compared to state-of-the-art WSSS techniques on two public histopathology datasets.
- Achieved higher mean Intersection over Union (mIoU) than fully supervised models.
- Showed comparable results in frequency-weighted Intersection over Union (fwIoU) to fully supervised models.
- Reduced annotation time from hours to minutes compared to manual labeling.
Conclusions:
- The MSRMMP framework effectively overcomes limitations of traditional WSSS methods, particularly for multi-target segmentation.
- Offers a significant reduction in annotation effort, making advanced image analysis more accessible.
- Presents a promising approach for accurate and efficient cancer diagnosis through automated histopathological image segmentation.
