SpikeDS uses dual sparsity to predict perineural invasion from 3D MRI

Perineural invasion (PNI) is a key prognostic factor in cholangiocarcinoma (CCA), but detecting it on 3D MRI remains difficult due to subtle, spatially sparse signals. Existing deep learning approaches are computationally expensive, limiting clinical adoption. To address this, a team of researchers introduced SpikeDS (Dual Sparsity Spikformer), a spiking neural network architecture that jointly leverages activation sparsity (from binary spike communication) and spatial sparsity (from pruning low-firing windows). The core innovation is Dual Sparsity Spiking Attention (DSSA), which combines Window-based Expert Mixture Spiking Attention (W-EMSA) — focusing computation on salient windows — and Cross-Window Spiking Self-Attention (CW-SSA), which allows pruned windows to contribute as key-value sources for global context.

Evaluated on a clinical cohort of 139 CCA patients using 5-fold cross-validation, SpikeDS achieved an AUC of 0.753 while consuming only 14.4 mJ of energy. This surpassed all baselines in both diagnostic performance and energy efficiency. The results demonstrate that dual sparsity offers a hardware-aware strategy for building efficient 3D spiking transformers without compromising accuracy. For clinicians and AI researchers, SpikeDS points toward a future where volumetric medical image analysis can be performed on edge devices or in resource-constrained settings, potentially enabling real-time PNI assessment at the bedside.