Comparison of eye position versus computer identified microcalcification clusters on mammograms

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Mammography is crucial for early breast cancer detection.
• Microcalcification clusters are key indicators but can be subtle.
• Computerized detection schemes aim to assist radiologists.

Purpose of the Study:

• To compare the accuracy of a computerized detection scheme with human observers (radiologists) in identifying microcalcification clusters on mammograms.
• To analyze discrepancies and overlaps in detection between the computer and radiologists.
• To investigate the role of eye-tracking data in understanding detection decisions.

Main Methods:

• Eighty digitized mammograms, half with subtle microcalcification clusters, were analyzed.
• A computerized detection scheme was applied.
• Six mammographers read the mammograms with recorded eye position.
• True and false positive locations were compared between computer and human readers.

Main Results:

• Computer achieved 83% true positives (0.5 false positives/image); radiologists ranged from 78-90% true positives (0.03-0.20 false positives/image).
• 90% of true clusters were detected by either the computer or human readers.
• Eye-tracking revealed differences in dwell time for true/false positives, correlating with calcification characteristics.

Conclusions:

• Computerized detection is a viable aid for mammographers, with performance comparable to human readers.
• Analysis of detection discrepancies can refine AI algorithms and improve understanding of human and computer false-positive decisions.
• Eye-tracking data provides insights into visual attention and feature interpretation in mammogram analysis.