Related Experiment Video
Updated: Jul 10, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
SC-SSL: Self-Correcting Collaborative and Contrastive Co-Training Model for Semi-Supervised Medical Image
IEEE Transactions on Medical Imaging
|November 23, 2023
Summary
This study introduces a novel self-correcting co-training scheme for semi-supervised learning in medical image segmentation. The method improves learning target quality for unlabeled data, enhancing segmentation accuracy and efficiency.
Area of Science:
- Medical Image Analysis
- Computer Vision
- Machine Learning
Background:
- Deep neural networks significantly advance medical image segmentation but require extensive labeled data, which is challenging to obtain.
- Semi-supervised learning (SSL) shows promise for medical image segmentation by leveraging unlabeled data, yet often overlooks the quality of learning targets derived from this data.
Purpose of the Study:
- To propose a novel self-correcting co-training scheme that generates improved learning targets for unlabeled data in medical image segmentation.
- To enhance the confidence and accuracy of segmentation models by refining the quality of targets used in semi-supervised learning.
Main Methods:
- A self-correcting module advances learning target generation as a learning task, increasing confidence for unannotated data.
- A structure constraint is imposed to enforce shape similarity between the improved learning target and collaborative network outputs.
- A pixel-wise contrastive learning loss is introduced to boost representation capacity, guided by the improved learning target for efficient exploration of unlabeled data with semantic context.
Main Results:
- The proposed method was extensively evaluated on four public datasets (ACDC, M&Ms, Pancreas-CT, Task_07 CT) against state-of-the-art semi-supervised approaches.
- Experimental results across varying labeled data ratios demonstrated the superiority of the proposed method over existing techniques.
- The approach effectively improves semi-supervised medical image segmentation performance.
Conclusions:
- The novel self-correcting co-training scheme significantly enhances semi-supervised medical image segmentation by improving learning target quality.
- The method's ability to generate more accurate targets and boost representation capacity leads to superior performance compared to current approaches.
- This work offers a more efficient and effective strategy for leveraging unlabeled data in medical image segmentation tasks.
