MRI-Based Deep-Learning Model for Distant Metastasis-Free Survival in Locoregionally Advanced Nasopharyngeal Carcinoma

Background Distant metastasis is the primary cause of treatment failure in locoregionally advanced nasopharyngeal carcinoma (LANPC). Purpose To develop a model to evaluate distant metastasis-free survival (DMFS) in LANPC and to explore the value of additional chemotherapy to concurrent chemoradiotherapy (CCRT) for different risk groups. Study Type Retrospective. Population In all, 233 patients with biopsy-confirmed nasopharyngeal carcinoma (NPC) from two hospitals. Field Strength 1.5T and 3T. Sequence Axial T-2-weighted (T-2-w) and contrast-enhanced T-1-weighted (CET1-w) images. Assessment Deep learning was used to build a model based on MRI images (including axial T-2-w and CET1-w images) and clinical variables. Hospital 1 patients were randomly divided into training (n =169) and validation (n =19) cohorts; Hospital 2 patients were assigned to a testing cohort (n =45). LANPC patients were divided into low- and high-risk groups according to their DMFS (P < 0.05). Kaplan-Meier survival analysis was performed to compare the DMFS of different risk groups and subgroup analysis was performed to compare patients treated with CCRT alone and treated with additional chemotherapy to CCRT in different risk groups, respectively. Statistical Tests Univariate analysis was performed to identify significant clinical variables. The area under the receiver operating characteristic (ROC) curve (AUC) was used to assess the model performance. Results Our deep-learning model integrating the deep-learning signature, node (N) stage (from TNM staging), plasma Epstein-Barr virus (EBV)-DNA, and treatment regimens yielded an AUC of 0.796 (95% confidence interval [CI]: 0.729-0.863), 0.795 (95% CI: 0.540-1.000), and 0.808 (95% CI: 0.654-0.962) in the training, internal validation, and external testing cohorts, respectively. Low-risk patients treated with CCRT alone had longer DMFS than patients treated with additional chemotherapy to CCRT (P < 0.05). Data Conclusion The proposed deep-learning model, based on MRI features and clinical variates, facilitated the prediction of DMFS in LANPC patients. Level of Evidence 3. Technical Efficacy Stage 4.

Automated summary

A deep-learning model based on MRI features and clinical variables was developed to predict distant metastasis-free survival in locoregionally advanced nasopharyngeal carcinoma patients. The model showed good performance in training, validation, and testing cohorts. Patients in low-risk groups treated with concurrent chemoradiotherapy alone had better outcomes compared to those who received additional chemotherapy.