TTIS: Training-free patch selection boosts whole slide image analysis
A new plug-and-play method selects only informative patches from gigapixel slides, no retraining needed.
In computational pathology, Whole Slide Image (WSI) analysis typically divides gigapixel slides into thousands of patches, which are then processed by Multiple Instance Learning (MIL) models. However, many patches contain redundant or non-informative tissue, wasting compute and potentially adding noise. Now, researchers Quoc Anh Nguyen, Sunhong Park, and Jin Tae Kwak propose Test-Time Instance Selection (TTIS), a training-free, plug-and-play framework that tackles this at inference time. TTIS automatically selects a compact yet representative subset of patches, discarding redundant ones without requiring any additional training or architectural changes.
The framework also includes a multi-view ensemble strategy that integrates multiple perspectives of tissue morphology, making the selection more robust. In extensive tests across several benchmark datasets, TTIS either improved or matched the performance of existing MIL models on classification and subtyping tasks. Because TTIS drops in seamlessly to current pipelines, it offers an accessible way to speed up whole slide analysis and reduce computational waste while preserving accuracy. Code has been released, making it easy for pathology AI teams to adopt.
- TTIS is a training-free, plug-and-play framework that selects compact, representative patches from whole slide images during inference.
- It integrates multi-view ensemble strategy to capture diverse tissue morphology and improve robustness without retraining.
- Across multiple benchmarks, TTIS improves or matches baseline MIL performance for cancer classification and subtyping tasks.
- Reduces computational overhead by discarding redundant, non-informative patches in gigapixel slides.
Why It Matters
Faster, cheaper whole slide analysis without retraining could accelerate AI adoption in pathology labs and improve cancer diagnostics.