Integrating Neuroscientific Priors into Spiking Neural Networks: ECSNN-SEG for Robust Brain ECS Segmentation from Low-SNR Cryo-Electron Microscopy Data

Medical image segmentation is crucial for computer-aided diagnosis, but segmenting complex structures like brain extracellular space (ECS) remains challenging due to low contrast, noise, and intricate morphology. Motivated by ECS’s role in neurophysiology, this study proposes ECSNN-SEG, a two-stage spiking neural network based on the biologically plausible ECS-LIF neuron model. It integrates a U-Net variant for coarse segmentation and an ECS-LIF-based refinement network to enhance accuracy and noise robustness. Evaluations on the cryo-EM ECSSeg dataset show ECSNN-SEG outperforms state-of-the-art methods in accuracy (ACC: 0.9834, F1: 0.8634) and robustness under Gaussian noise, highlighting the value of integrating neuroscientific priors into medical image analysis.