Explainable dual-task deep learning model for recurrence prediction and dose-sensitive region identification in HNC radiotherapy.

Abstract

Accurate prediction of distant metastasis (DM) and locoregional recurrence (LR) in head and neck cancer (HNC) patients following radiotherapy is critical for individualized treatment planning. This study aimed to develop and validate an explainable deep learning model integrating anatomical and dosimetric information for HNC DM and LR risk stratification. 237 HNC patients treated with definitive radiotherapy were collected and divided into training, internal, and external validation cohorts. A dual-task deep learning model based on 3D squeeze-and-excitation residual networks was constructed to predict DM and LR risk using CT images, 3D dose distributions, and gross tumor volume (GTV) masks. Model performance was evaluated by concordance index (C-index), time-dependent ROC, and decision curve analysis. Interpretability was enhanced using Grad-CAM and a novel Activation Volume Histogram (AVH) method to quantify attention patterns across spatial-dosimetric subregions. The model achieved C-indices of 0.92 and 0.74 for DM and LR, respectively, outperforming radiomic models. Grad-CAM visualizations revealed distinct activation patterns aligned with recurrence sites- lymphatic regions for DM and tumor zones for LR. AVH analysis identified GTV extended by 3mm receiving dose ≥ 65 Gy and GTV receiving dose ≥ 65 Gy as the most discriminative subregions, showing significant differences in activation distributions between high- and low-risk groups, particularly within the 0.4-0.5 intensity range. This dual-task model enables accurate DM risk prediction and showed potential for LR risk prediction in HNC, providing spatial-dosimetric information to support risk-adaptive radiotherapy. Further extensive multicenter validation is warranted to confirm its clinical applicability.