Diagnose Like A REAL Pathologist: An Uncertainty-Focused Approach for Trustworthy Multi-Resolution Multiple Instance Learning

TL;DR

UFC-MIL tackles the need for calibrated, trustworthy multi-resolution MIL in histopathology by modeling pathologist-like zooming behavior and uncertainty. It introduces patch-wise uncertainty loss, a Topological Neighbor Attention Module, and Uncertainty-Masked Cross-Attention, together with Sample-and-Resolution-Wise Label Smoothing for inference-free calibration. The approach enables end-to-end training across resolutions while delivering well-calibrated predictions with competitive accuracy on public datasets. These results advance clinically deployable MIL by combining uncertainty-aware reasoning with efficient multi-resolution analysis.

Abstract

With the increasing demand for histopathological specimen examination and diagnostic reporting, Multiple Instance Learning (MIL) has received heightened research focus as a viable solution for AI-centric diagnostic aid. Recently, to improve its performance and make it work more like a pathologist, several MIL approaches based on the use of multiple-resolution images have been proposed, delivering often higher performance than those that use single-resolution images. Despite impressive recent developments of multiple-resolution MIL, previous approaches only focus on improving performance, thereby lacking research on well-calibrated MIL that clinical experts can rely on for trustworthy diagnostic results. In this study, we propose Uncertainty-Focused Calibrated MIL (UFC-MIL), which more closely mimics the pathologists' examination behaviors while providing calibrated diagnostic predictions, using multiple images with different resolutions. UFC-MIL includes a novel patch-wise loss that learns the latent patterns of instances and expresses their uncertainty for classification. Also, the attention-based architecture with a neighbor patch aggregation module collects features for the classifier. In addition, aggregated predictions are calibrated through patch-level uncertainty without requiring multiple iterative inferences, which is a key practical advantage. Against challenging public datasets, UFC-MIL shows superior performance in model calibration while achieving classification accuracy comparable to that of state-of-the-art methods.

Figures (5)

• Figure 1: An illustration of the pathologists' observation pattern and the key mechanisms of UFC-MIL that reflect the pattern. (a) Pathologists begin observation at the coarsest resolution, identifying uncertain areas for further scrutiny. They zoom into these area to acquire additional information for diagnosis. (b) UFC-MIL, equipped with multi-resolution patches, focuses on sub-patches of those identified as uncertain at higher resolutions. Patch-level uncertainty at each resolution is then applied to calibration.
• Figure 2: Overview of UFC-MIL, which employs a top-down analysis from the coarsest $(r=1)$ to the finest $(r=R>1)$ resolution.
• Figure 3: Reliability diagrams on CAMELYON16. We plot histograms comparing uncalibrated models (a) with methods achieving the best ECE (b) for each. Analysis on DHMC and BCNB is presented in the supplementary material.
• Figure 4: Performance with all proposed methods is shown by a dashed line (i.e., UFC-MIL$^\bigstar$), with the difference from each ablation indicated above the bars.
• Figure 5: Illustration of attention map versus uncertainty map. In the attention map of HAG-MIL xiong2023diagnose, patches with low attention scores that were dropped during the zooming process are shown in grayscale, which the fine-grained model had no opportunity to observe. Additional cases are found in the supplementary material.