Development and validation of AI-based pre-screening of large bowel biopsies

Background Histopathological examination is a pivotal step in the diagnosis and treatment planning of many major diseases. With the aims of facilitating diagnostic decision-making and improving the use of pathologists’ time, we developed an AI-based pre-screening tool that analyses whole slide images (WSIs) of large bowel biopsies to identify normal, inflammatory, and neoplastic biopsies. Methods To learn the differential histological patterns from digitised WSIs of large bowel biopsy slides stained with Haematoxylin and Eosin (H&E), our proposed weakly supervised deep learning method uses only slide-level diagnostic labels and no detailed cell or region-level annotations. The proposed method was developed on an internal cohort of biopsy slides (n=5054) from a single laboratory labelled with corresponding diagnostic categories assigned by pathologists. Performance of the tool was evaluated on the internal development cohort (n=5054) in a cross-validation setting, and three external unseen cohorts (n=1536) for independent validation. Findings The proposed tool demonstrates high degree of accuracy to assist with the pre-screening of large bowel biopsies, being able to identify neoplastic biopsies (AUROC = 0·993), inflammatory biopsies (AUROC = 0·966) and all abnormal biopsies (AUROC = 0·979). On the three independent validation cohorts, it achieves AUROC values of 0·943, 0·958 and 0·964 for the detection of abnormal biopsies. Analysis of saliency maps confirms the representation of disease heterogeneity in model predictions and their association with relevant histological features. Interestingly, after examining diagnostic discrepancies between the proposed AI tool and original diagnostic labels, a panel of pathologists found that the proposed tool correctly identified a number of abnormal slides that had been initially reported as normal. Interpretations The proposed tool with its high sensitivity of detecting abnormal colorectal biopsies promises significant improvements in clinical workflow efficiency and assistance in diagnostic decision-making through pre-screening of normal biopsies. Funding Innovate UK on behalf of UK Research and Innovation.