SASHA: AI that analyzes gigapixel pathology slides with 80-90% less data

Deep neural networks are transforming histopathology, but whole-slide images (WSIs) are gigapixel-scale, making it computationally prohibitive to analyze every region at high resolution. Diagnostic labels are often available only at the slide level, and the informative tissue regions can occupy as little as 10% of the slide.

To tackle this, Tarun Gogisetty, Naman Malpani, Gugan Thoppe, and Sridharan Devarajan propose SASHA (Sequential Attention-based Sampling for Histopathological Analysis), a deep reinforcement learning framework accepted at NeurIPS 2025. SASHA first trains a lightweight, hierarchical attention-based multiple instance learning (MIL) model to learn which patches are likely to contain diagnostic features. Then, instead of processing the entire WSI at full resolution, SASHA uses a reinforcement learning agent to sequentially sample and zoom into only the most informative high-resolution patches — covering just 10–20% of the slide.

Experiments show that SASHA matches or exceeds the accuracy of methods that analyze the entire WSI at full resolution, while using a fraction of the computational and memory resources. It also significantly outperforms competing sparse sampling approaches. The authors propose SASHA as an intelligent sampling model for any medical imaging task involving exceptionally large images with sparsely informative features. Code is available on GitHub, enabling integration into clinical workflows. This efficiency gain could make AI-powered pathology more scalable and affordable for real-world hospitals.