Deep Multimodal Fusion Network for the Retinogeniculate Visual Pathway Segmentation

The segmentation of the retinogeniculate visual pathway (RGVP) is a significant quantitative tool for analyzing the anatomy and trajectory of individual RGVP. However, due to the complex structure and elongated morphology of the RGVP, it is difficult to accurately identify the intracranial course based on a single structural MRI or diffusion MRI. In this work, we propose a novel deep multimodal fusion network for the retinogeniculate pathway segmentation, which fuses the useful information from different modalities. Specifically, the proposed fusion model uses the supervised information generated by the spatial attention mechanism module to select useful information from the master modal and the assistant modal, where the T1 weighted (T1w) images that contributes most to the final segmentation result are used as the master modal and the FA images are used as the assistant modal. The results show that our method can achieve the best performance compared with the RGVP segmentation methods presented in the paper.