Artificial intelligence for ovarian cancer diagnosis from digital pathology slides

Digital pathology is a rapidly growing field, allowing for the development of assistive diagnostic tools. Many tools use artificial intelligence (AI) to automatically provide insights from whole slide images (WSIs), aiming to improve the accuracy, objectivity, and efficiency of the diagnostic process. Research has typically focused on the most common cancers, but less common cancers have received comparatively little attention. We focus on the histological subtyping of ovarian cancer, an essential diagnostic task for determining optimal treatments and prognoses. Through a systematic literature review, we find that previous research has been limited to model prototyping with small homogeneous datasets, with little focus on clinical utility. We perform the most thorough analyses of automated ovarian cancer histological subtyping to date, using the largest training dataset and evaluating models through cross-validation, hold-out testing, external validations, bootstrapping, and hypothesis testing. Analyses are based on attention- based multiple instance learning (ABMIL) with an ImageNet-pretrained ResNet50 backbone, a commonly used WSI classifier. The computational complexity of current AI models is a key limitation, with pathology labs typically not having sufficient hardware for model deployment. We propose an active tissue sampling technique and show that this approach can drastically reduce the computational burden of inference with minimal impact on diagnostic performance. ABMIL analyses tissue at only a single magnification, with high magnifications offering more cellular detail and low magnifications providing broader tissue context. We find that 10x magnification balances the cellular and histoarchitectural details to give the most accurate ovarian cancer subtyping performance, while drastically reducing the computational burden compared to the clinical standard 40x magnification. Recently, histopathology foundation models have promised to revolutionise diagnostic AI. We analyse 14 foundation models and confirm that they give significantly greater performance than previous feature extractors. In ABMIL, tissue patches are treated as independent of each other. We propose a multi-resolution patch graph network to better model spatial context and find this marginally improves performance. The optimal model, a combination of a foundation model and a graph, achieved five-class balanced accuracies of 88%, 99%, and 77% in three validation sets, where our baseline model achieved only 66%, 69%, and 52%, and individual pathologists achieved 74-91% concordance with similarly determined labels. This gives us confidence that AI models could have clinical utility, so future work should focus on the practicalities of implementation and real-world validation.