Comparative study of computational visual attention models on two-dimensional medical images

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Cognitive Science

Background:

• Computational models of visual attention are widely used in robotics but underexplored in medical imaging.
• Radiologists require efficient visual attention strategies for interpreting numerous clinical images under time constraints.
• Visual saliency maps, indicating regions of high contrast, are hypothesized to predict radiologist eye fixation points.

Purpose of the Study:

• To evaluate the performance of state-of-the-art visual saliency models on medical imaging modalities.
• To compare model performance against radiologists' actual eye movements.
• To determine the necessity of modality-specific tuning for saliency models in medical applications.

Main Methods:

• Comparison of 16 computational visual saliency models.
• Evaluation across three distinct medical imaging modalities.
• Validation of saliency maps against recorded radiologist eye-tracking data.

Main Results:

• Saliency models achieved competitive accuracy based on three evaluation metrics.
• Model performance ranking varied significantly across the evaluated medical imaging modalities.
• Model rankings on medical images differed substantially from rankings on natural image datasets (MIT300).

Conclusions:

• Current visual saliency models require modality-specific adaptation for effective use in medical imaging.
• Tuning saliency models for specific medical imaging types is crucial for applications like image compression and radiology education.
• Further research is needed to optimize computational visual attention for clinical diagnostic workflows.