Prediction of distant melanoma recurrence from primary tumor digital H&E images using deep learning.

Abstract

9577 Background: Patients with resected melanoma are at high risk for systemic recurrence. Quantifying the risk of recurrence can help identify the need for adjuvant immunotherapies, and accelerate the acquisition of survival statistics in clinical trials. Methods: 75 patients with stages I, II, and III melanoma seen at Columbia University Medical Center between 2001 and 2014 and The Icahn School of Medicine at Mount Sinai between 2000 and 2010 were included based on availability of tissue and 24 months of clinical follow-up. Images were scanned into tiff files (Aperio Biosystems). Deep neural net (DNN) architecture was designed consisting of convolutional and recurrent neural networks (CNN, RNN). Using QuPath open source software for nuclear segmentation and cell classification, we generated cell location, density, and clustering features to identify tissue areas for training of network parameters. DNN analyzes image and feature information locally within an H&E image, generating a prediction vote per region, and votes are averaged. Results: The model was validated on two independent external sets of stage I-III primary melanomas. Cohort 1 (Yale Medical Center), had n = 86 patients, of whom 49 were alive or had no evidence of disease at death (no DMR) and 37 died from melanoma. The second set, Cohort 2 (Geisnger Health Systems), had n = 29 patients, 15 without DMR and 14 with DMR. Prediction scores correlated with DMR status in both sets (AUC = 0.94 and 0.77 for Cohorts 1 and 2, respectively). A multivariable Cox proportional hazard model showed DNN recurrence prediction to be an independent prognostic factor for both Cohort 1 (HR = 2.54, 95% CI: 1.54-4.19, p = .0004***) and Cohort 2 (HR = 8.43, 95% CI: 2.58-27.51, p = .001**). Conclusions: We designed a DNN for quantitative prediction of melanoma recurrence from a H&E stained tissue. The prediction score warrants further study in larger patient cohorts and may constitute a novel digital pathology tool for the selection of melanoma patients for adjuvant immunotherapy. [Table: see text]