Discovering Salient Anatomical Landmarks by Predicting Human Gaze.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Anatomical landmarks are essential for medical imaging tasks but are typically predefined and manually annotated.
• Current methods rely on expert-defined landmarks and manual or machine learning-based annotation, which can be time-consuming and prone to error.

Purpose of the Study:

• To develop an automated method for discovering and localizing visually salient anatomical landmarks in medical images.
• To apply this method to fetal neurosonographic images and evaluate its utility in image registration.

Main Methods:

• Recorded fetal ultrasound scans with sonographer gaze-tracking to capture visual attention patterns.
• Trained a convolutional neural network (CNN) to predict gaze distribution (saliency maps) from ultrasound video frames.
• Extracted landmarks as local maxima of predicted saliency maps and matched them across images using feature clustering.

Main Results:

• The method successfully discovered and localized visually salient landmarks in fetal neurosonographic images.
• The discovered landmarks were effectively used in affine image registration.
• Achieved average landmark alignment errors ranging from 4.1% to 10.9% of the fetal head long axis length.

Conclusions:

• The proposed method automates the discovery and localization of anatomical landmarks, reducing reliance on manual annotation.
• Visually salient landmarks identified through AI and gaze-tracking are valuable for medical image registration.
• This approach holds potential for improving efficiency and accuracy in medical imaging analysis.