The rise of digital health has made vast patient data sets increasingly accessible. These datasets are highly valuable for the application of deep learning to identify existing and discover new digital biomarkers. These biomarkers are crucial for diagnosing diseases, predicting patient outcomes and predicting the efficacy of treatments. This talk will explore key challenges and prospects in the emerging area of computational pathology, with a focus on harnessing artificial intelligence and machine learning for early detection and tailored therapy of cancer.

The current melanoma staging system does not explain 26% of survival variance. Melanoma architectural morphology appears to be of greater significance than in other solid tumours, with Breslow thickness (subjective tumour depth as determined by the pathologist) remaining the strongest prognostic indicator. The application of convolutional neural networks whole slide images may reveal new insights into tumour morphology and patient prognosis.

This talk will outline the creation and development of a custom-designed 2-class (tumour/ not tumour) segmentation network. The CNN located invasive melanoma within primary resection specimens with a 97.64% sensitivity and 99.91% specificity per-pixel across international test sets from 5 different data sources. Furthermore, this CNN has been used to derive the first objective biomarkers in melanoma based on the tumour morphology, including the ‘Automated Breslow Thickness’ and the ‘Nodularity Index’.

Datasets containing whole slide images and molecular data have shown potential to advance diagnostic, prognostic and predictive understanding of cancer through multimodal deep learning. Addressing scarcity of multimodal data is a priority for this field to continue to progress. This talk will introduce the collaboration between Genomics England and the National Pathology Imaging Cooperative to enrich the 100,000 Genome Project Cohort by retrospectively digitising the slides for its cancer participants. An update on progress and next steps will be discussed, as well as an overview of early exemplar projects using the data generated so far.