Related Experiment Video
Updated: Jul 31, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
STAMP: A Self-training Student-Teacher Augmentation-Driven Meta Pseudo-Labeling Framework for 3D Cardiac MRI Image
S M Kamrul Hasan1, Cristian Linte2
1Center for Imaging Science, Rochester Institute of Technology, Rochester, NY, USA.
Medical Image Understanding and Analysis. Medical Image Understanding and Analysis (Conference)
|May 1, 2023
Summary
This study introduces STAMP, a semi-supervised learning framework that reduces pseudo-labeling bias in medical image segmentation. STAMP improves segmentation accuracy by adapting a teacher network and using data augmentation, outperforming existing methods.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Deep learning enhances medical image segmentation.
- Semi-supervised learning (SSL) leverages unlabeled data for improved performance.
- Pseudo-labeling bias is a key challenge in SSL for segmentation.
Purpose of the Study:
- To propose STAMP, a novel SSL framework to mitigate pseudo-labeling bias in medical image segmentation.
- To enhance model generalization and reduce error rates through adaptive pseudo-labeling and data augmentation.
Main Methods:
- Developed STAMP (Student-Teacher Augmentation-driven consistency regularization via Meta Pseudo-Labeling) framework.
- Employed a meta pseudo-labeling strategy where the Teacher network adapts based on Student network performance.
- Utilized strong and weak data augmentation policies for consistent probability distribution output.
Main Results:
- STAMP demonstrated effectiveness in segmenting left atrial cavities from GE-MR images.
- Achieved a 2.6% improvement in Dice and 4.4% improvement in Jaccard over state-of-the-art SSL methods.
- Outperformance was observed using only 10% labeled data, highlighting efficient unlabeled data exploitation.
Conclusions:
- STAMP effectively addresses pseudo-labeling bias in semi-supervised medical image segmentation.
- The framework shows significant potential for improving segmentation accuracy with limited labeled data.
- STAMP offers a robust solution for leveraging unlabeled data in medical imaging tasks.
