Self-Supervised Learning to Improve Topology-Optimized Axon Segmentation and Centerline Detection

Large-scale brain mapping requires preserving the topology of axonal structures to understand how neurons connect throughout the brain and intersect different brain regions. The ability to leverage unannotated data for algorithm development would mitigate laborious annotations by subject matter experts. In this work, we applied self-supervised training to a Residual 3D U-Net with an auxiliary classifier to predict the correct order of sliced and shuffled voxels samples of mouse brain data. Pretrained encoder weights from the auxiliary classifier were subsequently used to train a Residual 3D U-Net for axon segmentation and centerline detection with a topology-preserving loss function, soft centerline-Dice. We report improved feature space representation of axonal structures as well as improved evaluation metrics over prior methods. We observed that both the α and k hyperparameters of the topologically-aware loss function show sensitivity to the target task of axon segmentation and centerline detection.