Automated gastric cancer diagnosis on H&E-stained sections; training a classifier on a large scale with multiple instance machine learning

We present a system that detects cancer on slides of gastric tissue sections stained with hematoxylin and eosin (H&E). At its heart is a classifier trained using the semi-supervised multi-instance learning framework (MIL) where each tissue is represented by a set of regions-of-interest (ROI) and a single label. Such labels are readily obtained because pathologists diagnose each tissue independently as part of the normal clinical workflow. From a large dataset of over 26K gastric tissue sections from over 12K patients obtained from a clinical load spanning several months, we train a MIL classifier on a patient-level partition of the dataset (2/3 of the patients) and obtain a very high performance of 96% (AUC), tested on the remaining 1/3 never-seen before patients (over 8K tissues). We show this level of performance to match the more costly supervised approach where individual ROIs need to be labeled manually. The large amount of data used to train this system gives us confidence in its robustness and that it can be safely used in a clinical setting. We demonstrate how it can improve the clinical workflow when used for pre-screening or quality control. For pre-screening, the system can diagnose 47% of the tissues with a very low likelihood (< 1%) of missing cancers, thus halving the clinicians' caseload. For quality control, compared to random rechecking of 33% of the cases, the system achieves a three-fold increase in the likelihood of catching cancers missed by pathologists. The system is currently in regular use at independent pathology labs in Japan where it is used to double-check clinician's diagnoses. At the end of 2012 it will have analyzed over 80,000 slides of gastric and colorectal samples (200,000 tissues).