New AI method tackles gigapixel medical images by focusing on what matters

Pathologists examining tissue samples under a microscope can be waiting years for the benefits of fully automated analysis—largely because modern scanning equipment produces gigapixel-scale images that are computationally overwhelming to process in their entirety. A new approach, called SASHA (Sequential Attention-based Sampling for Histopathological Analysis), sidesteps that bottleneck by intelligently zeroing in on the most informative regions.

The core challenge#

Whole-slide images (WSIs) in histopathology contain billions of pixels, but the diagnostic information they carry is scattered thinly across them. Deep learning models trained to handle such data face two major headwinds: memory and compute constraints make it impractical to analyze every patch at full resolution, and training labels typically come only at the slide level—experts rarely annotate individual patches because it's labor-intensive and costly.

How SASHA works#

The method combines two key components. First, it learns a compact representation of the slide using a lightweight hierarchical attention-based multiple instance learning (MIL) model—a technique that treats the slide as a bag of patches and learns which ones matter most. Second, SASHA uses reinforcement learning to sample strategically, zooming into only 10–20% of high-resolution patches while discarding or examining at lower resolution the rest.

This dual approach lets the model focus computational budget where it counts, reducing both memory footprint and processing time without sacrificing diagnostic accuracy.

Performance and implications#

According to the research from arXiv, SASHA matches the diagnostic accuracy of state-of-the-art methods that laboriously analyze entire slides at high resolution—but at a fraction of the computational and memory cost. It also substantially outperforms existing sparse-sampling competitors.

The framework is positioned as a general solution for medical imaging tasks where images are exceptionally large yet contain sparse diagnostic signals. Implementation code is made available to the research community, potentially accelerating adoption in clinical and research pipelines.