This report introduces SMMILe, a novel computational pathology method designed to improve both whole-slide image classification and spatial quantification across various cancers. While traditional multiple-instance learning models often sacrifice spatial detail for global accuracy, SMMILe utilizes a superpatch-based approach to maintain high-resolution diagnostic maps without losing predictive power. The researchers mathematically demonstrate that instance-level aggregation overcomes the limitations of previous models, which often produce skewed or incomplete attention maps. By benchmarking the system across eight datasets and six cancer types, the study proves that SMMILe consistently matches or exceeds state-of-the-art performance in tasks like metastasis detection and tumor grading. Ultimately, this framework provides a robust tool for explainable AI, allowing pathologists to accurately identify and visualize specific tissue phenotypes within massive digital images.

References:

• Gao Z, Mao A, Dong Y, et al. SMMILe enables accurate spatial quantification in digital pathology using multiple-instance learning[J]. Nature Cancer, 2025: 1-17.