Human gaze-based dual teacher guidance learning for semi-supervised medical image segmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Medical image segmentation faces challenges due to limited labeled data, hindering model accuracy.
• Gaze data offers a cost-effective alternative to manual annotation for improving segmentation models.

Purpose of the Study:

• To address data scarcity and enhance network perception in semi-supervised medical image segmentation.
• To introduce a novel framework leveraging human gaze data within the mean-teacher architecture.

Main Methods:

• Proposed the Human Gaze-based Dual Teacher Guidance Learning (HG-DTGL) model, integrating human gaze as a guidance mechanism.
• Introduced GazeMix for dataset expansion and the Multi-scale Gaze Perception (MGP) module for enhanced feature extraction.
• Developed a Gaze Loss function to align model perception with human gaze patterns.

Main Results:

• HG-DTGL demonstrated superior performance in segmenting ten different organs/tissues across multiple medical imaging modalities.
• The model effectively expanded dataset scale and diversity, improving network perception capabilities.
• Extensive experiments confirmed the strong generalization ability of the proposed method.

Conclusions:

• Human gaze data significantly enhances semi-supervised medical image segmentation.
• The HG-DTGL model shows great potential for improving medical image analysis and clinical applications.