Lightweight hybrid CNN–SNN framework for ischemic stroke lesion segmentation

Accurate and effective segmentation of ischemic stroke lesions in Magnetic Resonance Imaging is essential for timely clinical intervention and treatment planning. Although Convolutional Neural Networks (CNNs) perform remarkably well in extracting spatial features, they may benefit from complementary temporal denoising mechanisms when handling the intrinsic noise present in stroke imaging. This paper presents a hybrid CNN–SNN framework that combines the spatial learning capability of CNNs with the spike-based temporal processing of Spiking Neural Networks (SNNs), incorporating Leaky Integrate-and-Fire neurons to suppress segmentation artifacts. On a patient-level, subject-wise split of the ISLES 2022 public training release (175/37/38 for train/validation/test), the proposed Hybrid CNN–SNN model achieves a Dice Similarity Coefficient of 0.599 and an Intersection over Union of 0.471 on full, unscrubbed 3D test volumes, with a 95th-percentile Hausdorff Distance (HD95) of 14.41 mm computed in true physical millimeters. The hybrid model is statistically indistinguishable from a CNN-only baseline on boundary localization (HD95, p = 0.38) while using 26% fewer parameters (92,661 vs. 125,557), and significantly outperforms an SNN-only baseline on Dice ( p < 0.0001). A grid search over post-processing configurations shows that thresholding and morphological refinement consistently reduce performance once evaluation is performed on full, unscrubbed patient volumes, and post-processing is therefore omitted from the final model. These findings position the hybrid CNN–SNN architecture as a compact, boundary-accurate alternative to conventional CNN segmentation, rather than a strict accuracy improvement over it, with potential relevance for resource-constrained clinical deployment.