DTC-WSI: Dynamic Token Compression for Whole-Slide Images

Abstract

Whole-slide images (WSIs) contain tens of thousands of heterogeneous patches, making transformer-based multiple-instance learning (MIL) computationally expensive due to quadratic attention costs and substantial redundancy in tissue morphology. Existing token-reduction approaches for WSI analysis rely primarily on pruning, which discards information early in training and destabilizes optimization under weak supervision. We propose Dynamic Token Compression for Whole-Slide Images (DTC-WSI), a token-efficient MIL framework that performs progressive, importance-aware WSI compression. DTC-WSI integrates a lightweight saliency network with a multi-stage token compressor that combines bipartite similarity matching and soft differentiable pruning to gradually eliminate redundant or non-diagnostic patches. During training, soft gates enable stable gradient flow, while inference employs deterministic compression for substantial acceleration. This curriculum-style compression preserves discriminative morphology and dramatically reduces computational burden. Across four WSI benchmarks (TCGA-NSCLC, TCGA-BRCA, TCGA-RCC, PANDA), DTC-WSI achieves 5–10$\times$ token reduction, up to 5.3$\times$ faster inference, and 20–40% lower memory usage, while improving MIL classification accuracy by 2–4% over state-of-the-art baselines. Our results demonstrate that dynamic token compression is a powerful and scalable alternative to pruning, enabling efficient transformer-based WSI analysis while improving accuracy.