UKSSL: Underlying Knowledge Based Semi-Supervised Learning for Medical Image Classification.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Medical image analysis benefits from deep learning but faces challenges with limited labeled data.
• Unlabeled medical images often outnumber labeled ones, hindering model training.
• Developing high-performance models with scarce labeled data is a critical research area.

Purpose of the Study:

• To introduce UKSSL, an underlying knowledge-based semi-supervised framework.
• To address the challenge of training effective medical image classification models with limited labeled data.
• To leverage unlabeled data for improved performance in medical image analysis.

Main Methods:

• UKSSL framework comprises two components: MedCLR for feature representation from unlabeled data and UKMLP for fine-tuning with labeled data.
• MedCLR extracts underlying knowledge from large unlabeled medical image datasets.
• UKMLP utilizes extracted features and limited labeled data for medical image classification.

Main Results:

• UKSSL achieved high performance on LC25000 and BCCD datasets using only 50% labeled data.
• On LC25000, UKSSL reported precision, recall, F1-score, and accuracy of 98.9%.
• On BCCD, UKSSL achieved 94.3% precision, 94.5% recall, 94.3% F1-score, and 94.1% accuracy, outperforming supervised methods with 100% labeled data.

Conclusions:

• UKSSL efficiently extracts valuable underlying knowledge from unlabeled medical image datasets.
• The framework demonstrates superior performance in medical image classification with limited labeled data.
• UKSSL offers a promising solution for data-scarce scenarios in medical deep learning.