UPMatch: Enhancing Semi-Supervised Medical Image Classification through Contrastive Learning with Unreliable Pseudo...

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning (SSL) is crucial for medical image analysis (MIA) due to limited labeled data.
• Existing SSL methods often underutilize unlabeled data by relying on strict confidence thresholds.
• Low inter-class distance and imbalanced categories in MIA exacerbate data underutilization.

Purpose of the Study:

• To develop a novel pseudo-labeling based SSL framework, UPmatch, to enhance information mining from ambiguous unlabeled samples in MIA.
• To improve model performance by effectively incorporating unreliable pseudo-label samples into the training process.
• To introduce an iterative informative sample selection strategy (ISSS) for contrastive learning.

Main Methods:

• Proposed UPmatch framework featuring a contrastive unreliable pseudo label learning module (CUPM).
• Implemented an informative sample selection strategy (ISSS) for iterative mini-batch sample selection.
• Evaluated the framework on TissueMNIST and ISIC2019 datasets.

Main Results:

• UPmatch demonstrated effectiveness in improving model performance on MIA tasks.
• The CUPM module successfully incorporated unreliable pseudo-label samples.
• The ISSS strategy enabled efficient selection of informative samples for contrastive learning.

Conclusions:

• The proposed UPmatch framework offers a significant advancement in SSL for MIA.
• Effectively utilizing ambiguous unlabeled data is key to improving model performance.
• UPmatch provides a promising approach for addressing data scarcity in medical imaging.