Dual-Uncertainty Guided Multimodal MRI-Based Visual Pathway Extraction

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Accurate extraction of the visual pathway (VP) from multimodal MRI is crucial for understanding brain structure and function.
• Current methods often require extensive manual annotation, posing a significant bottleneck in research and clinical applications.

Purpose of the Study:

• To develop a label-efficient approach for accurate visual pathway extraction from multimodal MRI.
• To minimize reliance on large labeled datasets while enhancing extraction performance.

Main Methods:

• Proposed a novel method incorporating a Modality-Relevant Feature Extraction Module (MRFEM) for T1-weighted and fractional anisotropy (FA) images.
• Implemented a mean-teacher model with dual uncertainty-aware ambiguity identification (DUAI) to improve VP extraction reliability.

Main Results:

• Demonstrated significant reduction in annotation efforts (at least one-third) compared to fully supervised methods on HCP and MDM datasets.
• Achieved superior extraction performance over six state-of-the-art semi-supervised methods.

Conclusions:

• The proposed label-efficient approach effectively reduces manual annotation burdens.
• Enhanced accuracy in multimodal MRI-based VP extraction facilitates improved analysis of complex brain structures.