Integrating computed tomography image features improves clinical prediction models for outcomes in nasopharyngeal carcinoma patients treated with (chemo)radiation.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiation Oncology, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands.
- Department of Radiation Oncology, Hunan Cancer Hospital, Xiangya School of Medicine, Central South University, Changsha, Hunan, China.
- Image Sciences Institute, University Medical Center Utrecht, Utrecht University, Utrecht, the Netherlands.
- Department of Radiation Oncology, China-Japan Friendship Hospital, Beijing, China.
- Hunan Institute of Schistosomiasis Control (The Third People's Hospital of Hunan Province), Yueyang, Hunan, China.
- Affiliated Nanhua Hospital, University of South China, Hengyang, Hunan, China.
Abstract
Clinical prognostic models for nasopharyngeal carcinoma (NPC) treated with intensity-modulated radiotherapy (IMRT) with or without chemotherapy remain insufficient to capture tumour heterogeneity. We investigated whether computed tomography (CT)-based signatures add prognostic value for overall survival, progression-free survival, local control and distant control in NPC patients. The study population consisted of 1360 patients with stage I-IVa NPC treated with (chemo)IMRT (2013-2017). Radiomic and deep-learning features were analysed with twelve clinical variables. Radiomic models were built using bootstrap resampling feature selection and multivariable Cox regression; deep-learning models used 3D ResNet-18 or DenseNet-121. Models were evaluated on an internal hold-out test set (<i>n</i> = 409; training set <i>n</i> = 951) with the concordance index and compared against clinical-only reference models. Decision curve analysis was used to assess clinical utility. Adding radiomic primary tumour features (Neighbouring Gray Tone Difference Matrix - coarseness) improved local control concordance index from 0.51 to 0.60 (<i>p</i> = 0.02). A DenseNet-121 combining clinical data with composite primary tumour and lymph node masks achieved the highest distant control (0.68 vs 0.66, <i>p</i> = 0.01). For overall survival and progression-free survival, the improvements were not significant. Decision curve analysis demonstrated net benefit of the DenseNet-121 distant control model over treat-all and treat-none strategies at threshold probabilities of 10-25%. Incorporating CT-based radiomic and deep-learning features into prognostic models significantly improved prediction of local and distant control in NPC, supporting their potential as imaging biomarkers for refined risk stratification.