DistAL: A Domain-Shift Active Learning Framework With Transferable Feature Learning for Lesion Detection.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Computer vision

Background:

• Deep learning models for medical imaging face performance degradation due to domain shifts across different institutions.
• Lesion detection is particularly vulnerable to domain shifts caused by variations in organs, diseases, imaging devices, and protocols.
• Manual annotation of medical images for training deep learning models is resource-intensive, time-consuming, and requires expert knowledge.

Purpose of the Study:

• To develop a cost-effective framework for improving deep learning model performance in medical image analysis across different domains.
• To reduce the burden of manual data annotation in medical imaging by intelligently selecting samples for labeling.
• To enhance the robustness and generalizability of lesion detection models despite domain variations.

Main Methods:

• Proposed a Domain-shift Active Learning (DistAL) framework combining active learning with domain-invariant feature learning.
• Employed contrastive-consistency training to learn discriminative and domain-invariant features.
• Introduced the RUDY (Representativeness, Uncertainty, and DiversitY) sample selection strategy for efficient and diverse data selection.

Main Results:

• DistAL achieved performance comparable to models trained on all target domain labels by annotating only 1.7% of target samples.
• The RUDY strategy effectively selected representative, uncertain, and diverse samples, mitigating domain shift challenges.
• Outperformed other active learning methods across five experiments on eight diverse datasets from different hospitals.

Conclusions:

• The proposed DistAL framework offers a viable solution for domain shift problems in medical image analysis, particularly for lesion detection.
• Combining active learning with domain-invariant feature learning significantly reduces annotation costs while maintaining high model accuracy.
• DistAL demonstrates superior performance and efficiency compared to existing methods, paving the way for more accessible AI in healthcare.