StaDis: Stability distance to detecting out-of-distribution data in computational pathology

Modern Computational pathology (CPath) models aim to alleviate the burden on pathologists. However, once deployed, these models may generate unreliable predictions when encountering data types not seen during training, potentially causing a trust crisis within the computational pathology community. Out-of-distribution (OOD) detection, acting as a safety measure before model deployment, demonstrates significant promise in ensuring the reliable use of models in real clinical application. However, most existing computational pathology models lack OOD detection mechanisms, and no OOD detection method is specifically designed for this field. In this paper, we propose a novel OOD detection approach called Stability Distance (StaDis), uniquely developed for CPath. StaDis measures the feature gap between an image and its perturbed counterpart. As a plug-and-play module, it requires no retraining and integrates seamlessly with any model. Additionally, for the first time, we explore OOD detection at the whole-slide image (WSI) level within the multiple instance learning (MIL) framework. Then, we design different pathological OOD detection benchmarks covering three real clinical scenarios: patch- and slide-level anomaly tissue detection, rare case mining, and frozen section (FS) detection. Finally, extensive comparative experiments are conducted on these pathological OOD benchmarks. In 38 experiments, our approach achieves SOTA performance in 23 cases and ranks second in 10 experiments. Especially, the AUROC results of StaDis with "Conch" as the backbone improve by 7.91% for patch-based anomaly tissue detection. Our code is available at https://github.com/zdipath/StaDis.