Consistency-based Semi-supervised Evidential Active Learning Framework for Robust Classification of Radiology Images

Area of Science:

• Medical Imaging AI
• Machine Learning for Healthcare
• Radiology Informatics

Background:

• Deep learning excels in radiology image classification but requires large, expert-annotated datasets.
• Semi-supervised learning and active learning can reduce annotation burden by utilizing unlabeled data.
• Combining these methods (SSAL) offers benefits but can be limited by unreliable uncertainty estimation and consistency enforcement.

Purpose of the Study:

• To propose Consistency-based Semi-supervised Evidential Active Learning (CSEAL) for robust radiology image classification.
• To leverage evidential learning for reliable uncertainty estimation and prioritized sampling in SSAL.
• To enhance the efficiency of clinical image annotation and model development workflows.

Main Methods:

• Developed evidential counterparts of semi-supervised methods (Pseudo-labelling, VAT, Mean Teacher, NoTeacher) within the CSEAL framework.
• Introduced Noise Robust-evidential NoTeacher using consensus and small-loss inclusion for handling noisy annotations.
• Evaluated CSEAL extensively on diverse X-ray, CT, and MRI datasets.

Main Results:

• CSEAL demonstrated substantial performance gains over existing SSAL baselines.
• The framework showed enhanced robustness in scenarios with noisy annotations.
• CSEAL-assisted annotation platforms achieved performance close to fully supervised learning with minimal labeled data.

Conclusions:

• CSEAL offers a principled and effective approach to semi-supervised evidential active learning for radiology.
• The method significantly improves model performance and annotation efficiency, even with imperfect labels.
• CSEAL presents a valuable tool for advancing clinical image annotation and AI model development.