Assessing saliency models of observers' visual attention on acquired facial differences

Area of Science:

• Computer vision
• Psychology
• Medical imaging

Background:

• Facial differences from head and neck cancer impact patients and families.
• Predicting visual attention to these differences can inform psychosocial interventions.
• Saliency models aim to predict where observers look.

Purpose of the Study:

• To evaluate existing saliency models' ability to predict visual attention to acquired facial differences.
• Focus on differences arising from head and neck cancer and its treatment.

Main Methods:

• Compared saliency maps from Graph-Based Visual Saliency (GBVS), Artificial Neural Network (ANN), and a face-specific model.
• Used eye-tracking data from lay observers viewing patient photographs.
• Employed linear mixed-effects modeling to analyze accuracy.

Main Results:

• GBVS model identified irrelevant salient regions (e.g., collars).
• ANN model inaccurately focused attention away from facial differences.
• Face-specific model showed higher accuracy but underestimated saliency of structural deviations.
• Facial difference location and interobserver variability significantly impacted model performance.

Conclusions:

• Existing saliency models are insufficient for predicting attention to acquired facial differences.
• Enhanced face-specific models are required for accurate prediction in head and neck cancer contexts.