Automated identification of Retinogeniculate Visual Pathway based on Multiscale Point Cloud Fusion Model Network (MSPF-Net)

The Retinogeniculate Visual Pathway (RGVP) is one of the primary neural pathways responsible for transmitting visual signals from the retina to the lateral geniculate nucleus. Accurate and complete identification of the RGVP is crucial for advancing our understanding of visual cognition and for improving neurosurgical planning. Recently, a number of automated deep learning-based approaches have been developed to overcome the limitations of traditional RGVP identification methods, which are often time-consuming and labor-intensive. However, most existing methods primarily depend on fiber geometry and ambiguous microstructural features, while overlooking inter-subject variability in RGVP trajectories - an issue we refer to as "inter-subject RGVP trajectory variability". In this study, we propose a novel Multiscale Point Cloud Fusion Network (MSPF-Net) for automated RGVP identification using diffusion magnetic resonance image tractography. To address the aforementioned challenges, MSPF-Net integrates three key modules: a Feature Concatenating Module to fuse positional and microstructural information at the point level, a Local Feature Extraction Module to capture local geometric patterns, and a Global Feature Extraction Module to learn global contextual representations. We evaluate MSPF-Net on both HCP and CHCP datasets to assess its generalizability and robustness. Furthermore, we identify and quantify similar streamline pairs in the CHCP dataset to validate the model's effectiveness in mitigating the "inter-subject RGVP trajectory variability" issue. Experimental results demonstrate that MSPF-Net achieves higher recognition accuracy than several state-of-the-art (SOTA) methods, and effectively mitigate "inter-subject RGVP trajectory variability" issue. Overall, this study highlights the strong potential of deep learning-based approaches for accurate and fully automated RGVP identification.