AI Support for Accelerating Histopathological Slide Examinations of Prostate Cancer in Clinical Studies

While studies in pathology are essential for the progress in the diagnostic and prognostic techniques in the field, pathologist time is becoming an increasingly scarce resource, and can indeed become the limiting factor in the feasibility of studies to be performed. In this work, we demonstrate how the Digital Pathology platform by CRS4, for supporting research studies in digital pathology, has been augmented by the addition of AI-based features to accelerate image examination to reduce the pathologist time required for clinical studies. The platform has been extended to provide computationally generated annotations and visual cues to help the pathologist prioritize high-interest image areas. The system includes an image annotation pipeline with DeepHealth-based deep learning models for tissue identification and prostate cancer identification. Annotations are viewed through the platform's virtual microscope and can be controlled interactively (e.g., thresholding, coloring). Moreover, the platform captures inference provenance information and archives it as RO-Crate artifacts containing data and metadata required for reproducibility. We evaluate the models and the inference pipeline, achieving AUC of 0.986 and 0.969 for tissue and cancer identification, respectively, and verifying linear dependence of execution speed on image tissue content. Finally, we describe the ongoing clinical validation of the contribution, including preliminary results, and discuss feedback from clinical professionals regarding the overall approach.