Publications Repository - Helmholtz-Zentrum Dresden-Rossendorf

For treatment individualisation of patients with locally advanced head and neck squamous cell carcinoma (HNSCC) treated with primary radiochemotherapy, we explored the capabilities of different deep learning approaches for predicting loco-regional tumour control (LRC) from treatment-planning computed tomography images. Based on multicentre cohorts for exploration (206 patients) and testing (85 patients), multiple deep learning approaches including extraction of deep features, transfer learning and complete training from scratch with 2D and 3D convolutional layers were assessed and compared to a clinical model including the tumour volume. Analyses were based on Cox proportional hazards regression and performance was assessed by the concordance index (C-index). While all 2D approaches showed similar or worse performance than the clinical model on the test cohort (C-index 0.39), 3D convolutional neural networks achieved improved discrimination (C-index 0.31) and patient stratification into high and low risk groups of tumour recurrence (p=0.001), in particular when using model ensembles instead of single models. Prospective validation of this result is planned.