Title: SpikeReg: Leveraging Spiking Neural Networks for Energy-Efficient 3D Deformable Medical Image Registration

Abstract:

Aligning anatomical structures in 3D medical images through deformable registration is a process that remains heavily computationally intensive. While Spiking Neural Networks (SNNs) provide a mechanism for sparse, event-driven computation, their application to this specific task has not been thoroughly investigated. To address this gap, we present SpikeReg, a novel spiking U-Net architecture designed for 3D brain MRI registration.

The model is derived from an analog ANN registration teacher via layer-wise weight transfer and activation-percentile threshold calibration. It is subsequently refined using a surrogate-gradient objective function that integrates diffusion regularization, local cross-correlation, and spike-rate sparsity.

Evaluations on the OASIS Learn2Reg validation split, comprising 19 image pairs, demonstrate that SpikeReg achieves a Dice score of $0.7474 \pm 0.032$. This performance is statistically indistinguishable from its ANN teacher baseline ($0.7480 \pm 0.037$, $p = 0.67$). Notably, the model operates at a mean spike rate of $12.8\%$, yielding a projected $55.5\times$ reduction in arithmetic energy consumption compared to the dense-ANN baseline, as estimated by an event-sparse SynOps/MAC proxy.

Our study also highlights two critical negative findings: displacement distillation from the ANN teacher degrades performance, and ANN teachers optimized with a label-Dice loss are unable to transfer effectively through rate-code conversion. Collectively, these findings confirm that dense geometric prediction is feasible within sparse, event-driven computational frameworks, paving the way for neuromorphic approaches in medical image registration.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC