Predicting Clinician Fixations on Glaucoma OCT Reports via CNN-Based Saliency Prediction Methods

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Physician interpretation of optical coherence tomography (OCT) reports is crucial for diagnosing eye conditions like glaucoma.
• Understanding visual attention patterns during OCT report analysis can enhance diagnostic accuracy and training.
• Current methods for predicting physician focus on medical reports are limited.

Purpose of the Study:

• To predict physician fixations on ophthalmology OCT reports using Convolutional Neural Network (CNN) based saliency prediction.
• To develop a tool aiding the education of ophthalmologists and ophthalmologists-in-training by identifying key areas in OCT reports.
• To evaluate the accuracy of CNN-based saliency prediction models in the context of ophthalmology diagnostics.

Main Methods:

• Fifteen ophthalmologists evaluated 20 OCT reports each, assessing glaucoma likelihood.
• Eye-tracking data was collected using a Pupil Labs Core eye-tracker.
• Fixation heat maps were generated from eye-tracking data to analyze visual attention.

Main Results:

• A traditional saliency mapping model achieved a correlation coefficient (CC) of 0.208.
• Normalized Scanpath Saliency (NSS), Kullback-Leibler divergence (KLD), and Structural Similarity Index (SSIM) values were 0.8172, 2.573, and 0.169, respectively.
• The TranSalNet model demonstrated reasonable accuracy in predicting fixations on specific OCT report regions.

Conclusions:

• The TranSalNet model shows potential for predicting physician attention on OCT reports.
• Further research with increased and improved data is necessary to enhance model performance.
• Future work will focus on expanding data collection, refining data quality, and optimizing the model architecture.