Pathologists spend years learning where to look on a slide. Most AI models brute-force the entire thing. SASHA, a new framework from the Indian Institute of Science, Bangalore, does neither—it learns to hunt for cancer the way a doctor would.
The method, accepted at NeurIPS 2025 and published on arXiv in February 2026, tackles the gigapixel problem: whole-slide images are so massive that even powerful GPUs choke on them. SASHA uses a lightweight attention-based multiple instance learning (MIL) model to scout the slide for suspicious regions, then deploys a deep reinforcement learning (RL) agent that sequentially zooms into just 10–20% of the high-resolution patches.
Here's the counter-intuitive part: despite seeing only a fraction of the data, SASHA matches state-of-the-art methods that scan every pixel at full resolution. It significantly outperforms other sparse sampling techniques. That's like reading only the first and last chapters of a mystery novel and still solving the crime.
But here's the twist—and it's juicy. That RL agent requires training itself. The 'efficient' model has a higher upfront computational cost. The paper argues that, amortized over thousands of slides, the total is lower. But if you're a lab processing just a few dozen slides a day, the math might not work. The efficiency paradox: cheaper per slide, but you pay to learn how to be cheap.
Who wins? Pathologists in resource-constrained settings, where running a full WSI model on a high-end GPU is impossible. AI companies like Aiforia, which have already integrated similar attention-based grading for mouse lung tumor studies. Who loses? Vendors of brute-force whole-slide models that require full resolution—their cost advantage evaporates when a lightweight method matches their accuracy.
The lack of public controversy is suspicious. No one is arguing. That usually means either the method is universally loved, or the debate hasn't started yet. I suspect the latter. Reinforcement learning in medical imaging is rare, and RL agents can learn shortcuts—focusing on scanner artifacts rather than real pathology. Until SASHA passes a prospective clinical study on diverse slides with real-world artifacts (folds, air bubbles, variable staining), the hype is just academic.
Trends suggest SASHA could become the default baseline for attention-based WSI analysis—if the author releases code under a permissive license. If not, re-implementations might differ subtly and the method may never standardize. The real test? Extending SASHA to radiology or 3D medical imaging. The sequential attention + RL paradigm is modality-agnostic. That's where the next battle will be fought.
For now, SASHA is the smartest pathologist you've never met. It doesn't need to see everything—it just needs to know where to look.