Related Experiment Video
Updated: Aug 10, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
FaxMatch: Multi-Curriculum Pseudo-Labeling for semi-supervised medical image classification
Zhen Peng1,2, Dezhi Zhang3, Shengwei Tian1,2
1College of Software, Xinjiang University, Urumqi, China.
Medical Physics
|February 13, 2023
Summary
This study introduces Multi-Curriculum Pseudo-Labeling (MCPL) for semi-supervised learning in medical imaging. MCPL dynamically adjusts pseudo-labeling thresholds per class, outperforming fixed methods and improving model accuracy with limited labeled data.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Semi-supervised learning (SSL) leverages unlabeled data to enhance model performance, crucial for medical imaging tasks.
- Pseudo-labeling is a common SSL technique, but fixed thresholds for selecting pseudo-labels are suboptimal, especially with imbalanced medical data.
- Data imbalance in medical imaging poses challenges for traditional pseudo-labeling methods that ignore class-specific learning statuses.
Purpose of the Study:
- To address the limitations of fixed thresholds in pseudo-labeling for imbalanced medical image datasets.
- To develop an adaptive algorithm that accounts for varying class learning difficulties in semi-supervised medical image classification.
- To improve the efficiency and accuracy of semi-supervised learning models in medical imaging by optimizing pseudo-label generation.
Main Methods:
- Proposing Multi-Curriculum Pseudo-Labeling (MCPL) to dynamically adjust class-specific thresholds based on model learning status per epoch.
- Integrating MCPL with the FixMatch framework to create an improved semi-supervised learning algorithm for medical image classification, termed FaxMatch.
- Employing label smoothing (LS) to generate soft labels (SL) for pseudo-labels, mitigating the impact of potentially incorrect pseudo-labels.
Main Results:
- MCPL-based FaxMatch demonstrated superior performance compared to fully supervised baselines on ISIC 2018 and COVID-CT datasets.
- The proposed method achieved better results than existing state-of-the-art semi-supervised learning techniques.
- Extensive experiments confirmed the effectiveness of the adaptive thresholding strategy in improving classification accuracy.
Conclusions:
- MCPL offers an effective solution for semi-supervised medical image classification, particularly in scenarios with limited labeled data.
- The developed framework reduces dependency on large labeled datasets, thereby decreasing manual annotation efforts.
- This approach enhances the practical applicability of SSL in medical imaging by improving model robustness and performance.
