Chest X-Ray Visual Saliency Modeling: Eye-Tracking Dataset and Saliency Prediction Model.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Radiology and diagnostic imaging

Background:

• Radiologists' eye movements during medical image interpretation offer insights into their diagnostic decision-making processes.
• Eye-tracking data can be modeled to identify clinically relevant regions within medical images.
• This information holds potential for integration into artificial intelligence (AI) systems for automated medical image diagnosis.

Purpose of the Study:

• To establish a comprehensive chest X-ray (CXR) visual saliency benchmark using a large-scale eye-tracking study.
• To quantify the reliability and clinical relevance of saliency maps (SMs) derived from CXR images.
• To develop and validate a novel saliency prediction model (CXRSalNet) for enhancing AI diagnostic capabilities.

Main Methods:

• Conducted a large-scale eye-tracking study with 13 radiologists interpreting 191 CXR images.
• Analyzed gaze data to generate and evaluate saliency maps (SMs) for clinical relevance and reliability.
• Developed CXRSalNet, a saliency prediction model leveraging radiologists' gaze information for training on unlabeled CXR images.

Main Results:

• Established a best-of-its-kind CXR visual saliency benchmark.
• Quantified the reliability and clinical relevance of saliency maps in CXR interpretation.
• Demonstrated that CXRSalNet effectively utilizes gaze data to enhance AI diagnostic imaging system performance.

Conclusions:

• Radiologists' eye movements provide valuable data for understanding diagnostic processes in medical imaging.
• The developed CXRSalNet model effectively leverages gaze data to improve AI-driven diagnostic accuracy, particularly in data-scarce scenarios.
• This approach shows promise for enhancing the performance of AI systems in medical image interpretation.