<scp>AI</scp> ‐derived prognostic biomarkers from melanoma whole slide image segmentation: an initial discovery and assessment

The current melanoma staging system predicts 74% of the variance in survival, with prognostic biomarkers subject to high levels of inter-observer variation. This work assesses whether a previously developed convolutional neural network (CNN) for invasive melanoma segmentation in whole slide images (WSIs) may reveal new insights into melanoma morphology and patient prognosis. This paper uses Cox proportional multivariate regression analyses to evaluate the ability of the CNN outputs to predict patient survival across 745 WSIs from 5 data sources. Five objective histomorphological parameters of tumour size and shape that are independently associated with overall and melanoma-specific survival were created from the CNN: tumour area(log) (HR 1.48 CI 1.30-1.68, p < 0.001), tumour perimeter(log) (HR 1.86 CI 1.48-2.32, p < 0.001), major axis length(log) (HR 1.88 CI 1.42-2.48, p < 0.001), Nodularity Index(log) (HR 1.77 CI 1.28-2.43, p < 0.001) and digital Breslow thickness(log) (HR 2.04, CI 1.63-2.54, p < 0.001). These results indicate that melanoma segmentation of the entire lesion within a WSI may be used to predict patient outcome. Moreover, this technology can be used to make new morphological discoveries to provide information not currently contained within our staging system (e.g. Nodularity Index), as well as provide objectivity and automation of current biomarkers (e.g. digital Breslow thickness). Further work is required to validate this initial discovery and evaluation.