Eye Tracking-Enhanced Deep Learning for Medical Image Analysis: A Systematic Review on Data Efficiency,...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Human-Computer Interaction

Background:

• Deep learning (DL) excels in medical image analysis (MIA) but faces challenges with limited data and lack of transparency.
• Integrating DL into clinical practice is hindered by data scarcity and the 'black-box' nature of AI decisions.
• Physician trust and accountability are crucial for AI adoption in healthcare.

Purpose of the Study:

• To review state-of-the-art strategies for integrating eye tracking (ET) with deep learning (DL) in medical image analysis.
• To propose a unified framework for ET-DL integration that enhances data efficiency, model interpretability, and multimodal alignment.
• To address key implementation barriers for clinician-centered AI systems in healthcare.

Main Methods:

• Systematic literature review following PRISMA guidelines.
• Analysis of existing research on eye tracking (ET) for supervising and interpreting deep learning (DL) models.
• Development of a conceptual framework for ET-DL integration.

Main Results:

• Eye tracking (ET) provides supervisory signals to improve DL model accuracy and feature recognition, especially with limited data.
• ET facilitates comparison between machine and human attention, enhancing DL model interpretability and clinician trust.
• The proposed framework positions ET as a tool for data efficiency, interpretability validation, and multimodal alignment.

Conclusions:

• ET-DL integration offers a promising solution to enhance AI in medical image analysis, addressing data limitations and interpretability issues.
• A unified framework for ET-DL can lead to clinician-centered AI systems with verifiable reasoning and seamless workflow integration.
• This approach paves the way for the future clinical deployment of trustworthy and intelligible AI in healthcare.