Static and dynamic edge convolution for PointNet-based white matter fiber segmentation

The definition of white matter fiber tracts is a crucial step in fiber tract-based diffusion MRI analysis frameworks. Deep learning-based methods for automated fiber tract segmentation have shown reasonable success, though often needing time-intensive manual corrections. In this paper, we presented a novel fiber tract segmentation method that extends the established PointNet approach with static and dynamic edge convolution layers. Dynamic edge convolution (DEC) layers construct the graph dynamically based on the evolving feature space using k-NN, offering greater flexibility for tasks where the relationships between nodes are not fixed. Combined with an advanced data augmentation setup, and a confidence branch in the model, the proposed method performs at a high level, significantly outperforming the original PointNet and DEC methods.