Introduction
In digital pathology, Whole Slide Images (WSIs) present a unique computational challenge. These images are massive, often containing billions of pixels, yet the diagnostic signal—such as a small cluster of tumor cells—may only occupy a tiny fraction of the total area. This "needle in a haystack" scenario is typically addressed using Multiple Instance Learning (MIL), where a slide is treated as a "bag" of image patches (instances).
However, existing MIL frameworks face three significant hurdles:
- The Dilution Problem (Distribution): The overwhelming number of healthy (negative) patches can drown out the signal from a few cancerous (positive) ones during feature aggregation.
- The Context Gap (Correlation): Many models treat patches as independent, ignoring the spatial architecture and regional context essential for diagnosis.
- The Zoom Factor (Magnification): Pathologists naturally transition between different magnifications (e.g., 5x to 20x). Most AI models struggle to effectively fuse information across these scales.
To address these, we present MILBooster, a novel dual-scale, multi-stage framework designed to optimize WSI classification through improved distribution filtering, spatial correlation modeling, and multi-magnification fusion.
Methodology
MILBooster introduces three core components to refine the MIL pipeline: the Bag Filter, the PiceBlock, and the Scale Mixer.
The Bag Filter
The Bag Filter is a "plug-and-play" pre-processing module designed to solve the dilution problem. It maintains a bank of "true negative" features from the training set. By calculating the Mahalanobis distance between new patches and these known negatives, the filter identifies and removes redundant, "obviously healthy" instances. This process increases the concentration of potential positive instances, clarifying the decision boundary for the final classifier.
PiceBlock (Part Interact with Entirety Block)
Standard Transformers are often computationally prohibitive for the thousands of patches in a WSI. The PiceBlock addresses this by using a window-based approach. It models relationships within local windows and then merges these windows into "region features" to capture global context. This hierarchical approach mimics human visual perception, moving away from the "bag of words" assumption to respect the actual tissue architecture.
Scale Mixer and Overall Architecture
The overall PiceFormer architecture utilizes a dual-branch system to process different magnifications (e.g., 10x and 20x) in parallel. The Scale Mixer facilitates continuous communication between these branches throughout the learning process, rather than simple late-stage concatenation.
Experimental Results
MILBooster was evaluated across four clinical tasks using three diverse datasets: Camelyon16 (breast cancer metastasis), TCGA Lung Cancer (subtyping), and an in-house Cervical Cancer dataset (metastasis and prognosis).
- State-of-the-Art Performance: MILBooster consistently outperformed previous benchmarks like DSMIL and TransMIL. On Camelyon16, it achieved an AUC of 0.9427. On TCGA Lung Cancer, it reached a near-perfect AUC of 0.9863.
- Challenging Tasks: In cervical cancer prognosis—a notoriously difficult task—it achieved an AUC of 0.8036, a 1.2% improvement over the next best method.
- Efficiency: Despite its complexity, the PiceBlock is more efficient than a standard ViT-B/16, requiring fewer FLOPs (48.7G vs 55.4G) while delivering higher accuracy.
Discussion and Conclusion
The success of MILBooster demonstrates that "less is more" in WSI classification; removing up to 60% of redundant instances via the Bag Filter actually improves model performance. By combining this filtering with spatially-aware PiceBlocks and multi-scale fusion, the framework provides a more robust tool for digital pathology.
Future work will focus on reducing the dependency on a pre-constructed negative feature bank and extending the framework to support more than two simultaneous magnifications. Nonetheless, the current results suggest that MILBooster could significantly improve diagnostic accuracy in clinical settings, potentially leading to more reliable identifications for patients.
Summary updated