Deep learning reveals prognostic information beyond AJCC stage in routine histopathology of early-stage melanoma: the population-based D-ESMEL study

Of all melanoma patients who develop distant metastasis over time, 41% were initially diagnosed with stage I/II. More accurate prognostic models are needed to select patients for potential (neo)adjuvant therapy. We hypothesise that histopathological features in whole slide images (WSIs) from primary early-stage melanomas provide additional information beyond American Joint Committee on Cancer (AJCC) stage. We utilised data from the population-based Dutch Early-Stage Melanoma (D-ESMEL) study. This included a discovery set (n=328) of case-control sets matched on follow-up time and known prognostic features (i.e. Breslow thickness, ulceration, age, and sex) and a validation cohort based on >5,500 patients following a nested case-control (NCC) design with stage-matched controls (n=266). We developed an end-to-end deep learning workflow to automatically extract relevant patches from WSIs and predict patient-level metastatic risk based on histopathological features independent of AJCC stage. The discovery set served as a training set for the workflow using 10-fold cross-validation. The prognostic performance of this model and its independence from clinical risk factors was evaluated in the validation cohort using odds ratios (OR) and the weighted C-index. We achieved a balanced accuracy of 0.88 (mean: 0.81; 95% CI: 0.77-0.85) in the best validation fold of the discovery set. In the validation cohort, metastatic risk predictions from the deep learning model were an independent prognostic factor (adjusted OR: 1.97; p<0.01). Combining the deep learning predictions (C-index: 0.86) with AJCC stage (C-index: 0.80) yielded the highest C-index of 0.87, significantly improving the C-index compared to AJCC stage alone (p<0.001). Our histopathology-based deep learning model was rigorously evaluated in a population-based cohort and demonstrated added prognostic value beyond AJCC stage. Further investigations of the underlying biological features could improve clinical decision making and patient outcomes.