Spectral Multiple-Instance Learning for Efficient Gigapixel Image Analysis

ICLR 2026 Conference Submission25353 Authors

Keywords: Multiple-Instance Learning, Spectral Methods, Whole Slide Images

Abstract: With ongoing advances in imaging technology, gigapixel images are now widely utilized in both scientific research and industrial applications. However, their extremely large scale presents significant challenges for conventional deep learning workflows. A common approach involves partitioning the image into thousands of smaller patches, processing each patch independently, and aggregating the representations using a Multiple-Instance Learning (MIL) framework. Because the label of a gigapixel image often depends on a small subset of informative regions, identifying these key patches is essential. However, MIL faces a persistent multi-resolution dilemma: low-magnification views offer global contextual information but fail to capture fine-grained details, whereas high-magnification views retain these details at a substantial computational cost. We introduce Multi-Instance Learning with Spectral Methods (SpecMIL), which addresses this challenge by capturing high-frequency features at low magnification and preserving geometric relationships across scales using graph spectral theory. SpecMIL exploits spectral features that remain informative even after down-sampling, guiding selective high-resolution "zoom-in" only where necessary. Experiments on various whole slide image benchmarks (e.g., tumor subtyping, grading, and metastasis detection) demonstrate that spectral approaches offer a highly effective and efficient solution for gigapixel image analysis.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 25353