RGVPSeg: multimodal information fusion network for retinogeniculate visual pathway segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Accurate segmentation of the retinogeniculate visual pathway (RGVP) is crucial for quantitative anatomical analysis.
• Multimodal learning using structural MRI (sMRI) and diffusion MRI (dMRI) shows promise for RGVP segmentation.
• Challenges include the complex skull base and RGVP's slender morphology, hindering effective multimodal information fusion.

Purpose of the Study:

• To develop a multimodal information fusion network (RGVPSeg) for enhanced RGVP segmentation.
• To optimize and select complementary information from T1-weighted (T1w), fractional anisotropy (FA), and fiber orientation distribution function (fODF) peak images.
• To enable mutual supervision between modalities for improved segmentation accuracy.

Main Methods:

• Proposed a multimodal information fusion network (RGVPSeg) incorporating a supervised master-assistant cross-modal learning framework.
• Integrated T1w, FA, and fODF peak data for segmentation.
• Applied a cross-modal learning framework between encoder layers to leverage complementary information.

Main Results:

• RGVP segmentation was performed on datasets from the Human Connectome Project (HCP) and Multi-shell Diffusion MRI (MDM).
• Experimental results demonstrated promising segmentation accuracy.
• The proposed RGVPSeg framework outperformed existing methods in feasibility and performance.

Conclusions:

• The RGVPSeg framework effectively segments the retinogeniculate visual pathway using multimodal MRI data.
• The proposed cross-modal learning approach enhances the utilization of complementary information from different MRI modalities.
• RGVP segmentation accuracy is improved, offering a valuable tool for quantitative analysis of visual pathway anatomy.