Automatic Lung Nodule Detection Combined With Gaze Information Improves Radiologists' Screening Performance.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Early lung cancer diagnosis using computed tomography (CT) is crucial for reducing mortality.
• Identifying lung nodules in CT scans is complex and impacts screening program effectiveness.
• Computer-aided detection (CAD) systems can assist radiologists but may introduce bias and time constraints.

Purpose of the Study:

• To evaluate the utility of gaze information for integrating CAD systems into clinical practice for lung nodule detection.
• To assess how eye-tracking data can optimize the use of automated detection tools.

Main Methods:

• Four radiologists annotated 20 CT scans from a public dataset while their eye movements were tracked.
• An automatic lung nodule detection system was developed and evaluated.
• Radiologist search patterns and fixation durations were analyzed in relation to detection errors.

Main Results:

• Radiologists exhibited consistent search routines and shorter fixation times in areas with missed nodules.
• The system achieved a detection sensitivity of 0.69, comparable to individual radiologists (0.67±0.07).
• Combining one radiologist's annotations with the CAD system significantly improved detection sensitivity, matching the performance of two radiologists. Filtering CAD candidates based on low fixation regions enhanced sensitivity without increasing false positives.

Conclusions:

• Gaze information can be effectively leveraged to integrate CAD systems into clinical workflows for improved lung nodule detection.
• Combining radiologist expertise with AI-driven candidate filtering, guided by eye-tracking data, offers a promising strategy for enhancing lung cancer screening accuracy.
• This approach has the potential to improve the efficiency and effectiveness of lung cancer screening programs.