Explainable dual-task deep learning model for recurrence prediction and dose-sensitive region identification in HNC
Huali Li1, Jiajun Cai2, Xuanru Zhou3
1School of Biomedical Engineering, Southern Medical University, Guangzhou, Guangdong, China.
International Journal of Radiation Oncology, Biology, Physics
|August 13, 2026
Summary
This study developed an explainable deep learning model for head and neck cancer (HNC) patients, accurately predicting distant metastasis (DM) and showing potential for locoregional recurrence (LR) risk. The model integrates imaging and dose data to aid personalized radiotherapy.
Area of Science:
- Oncology
- Radiotherapy
- Artificial Intelligence
- Medical Imaging
Background:
- Accurate prediction of distant metastasis (DM) and locoregional recurrence (LR) is crucial for personalized treatment in head and neck cancer (HNC) patients undergoing radiotherapy.
- Current methods may lack the precision needed for effective risk stratification and individualized treatment planning.
Purpose of the Study:
- To develop and validate an explainable deep learning model for predicting DM and LR risk in HNC patients.
- To integrate anatomical and dosimetric information for improved HNC risk stratification.
- To provide interpretable insights into the model's predictions using novel visualization techniques.
Main Methods:
- A dual-task deep learning model using 3D squeeze-and-excitation residual networks was developed.
- The model utilized CT images, 3D dose distributions, and gross tumor volume (GTV) masks from 237 HNC patients.
- Model performance was assessed using concordance index (C-index) and ROC analysis, with interpretability enhanced by Grad-CAM and Activation Volume Histogram (AVH).
Main Results:
- The model achieved high performance with C-indices of 0.92 for DM and 0.74 for LR, outperforming radiomic models.
- Explainability methods revealed distinct spatial-dosimetric patterns associated with DM (lymphatic regions) and LR (tumor zones).
- AVH analysis identified specific dose-volume regions within the GTV as highly discriminative for risk stratification.
Conclusions:
- The developed dual-task deep learning model accurately predicts DM risk and shows promise for LR risk prediction in HNC.
- The model provides valuable spatial-dosimetric insights, supporting risk-adaptive radiotherapy strategies.
- Further multicenter validation is recommended to confirm clinical applicability.