MRI-Based Back Propagation Neural Network Model as a Powerful Tool for Predicting the Response to Induction Chemotherapy in Locoregionally Advanced Nasopharyngeal Carcinoma.

Background: Pretreatment individualized assessment of tumor response to induction chemotherapy (ICT) is a need in locoregionally advanced nasopharyngeal carcinoma (LANPC). Imaging method plays vital role in tumor response assessment. However, powerful imaging method for ICT response prediction in LANPC is insufficient.

Purpose: To establish a robust model for predicting response to ICT in LANPC by comparing the performance of back propagation neural network (BPNN) model with logistic regression model.

Study Type: Retrospective.

Population: A total of 286 LANPC patients were assigned to training (N = 200, 43.8 ± 10.9 years, 152 male) and testing (N = 86, 43.5 ± 11.3 years, 57 male) cohorts.

Field Strength/sequence: T -weighted imaging, contrast enhanced-T -weighted imaging using fast spin echo sequences at 1.5 T scanner.

Assessment: Predictive clinical factors were selected by univariate and multivariate logistic models. Radiomic features were screened by interclass correlation coefficient, single-factor analysis, and the least absolute shrinkage selection operator (LASSO). Four models based on clinical factors (Model ), radiomics features (Model ), and clinical factors + radiomics signatures using logistic (Model ), and BPNN (Model ) methods were established, and model performances were compared.

Statistical Tests: Student's t-test, Mann-Whitney U-test, and Chi-square test or Fisher's exact test were used for comparison analysis. The performance of models was assessed by area under the receiver operating characteristic (ROC) curve (AUC) and Delong test. P < 0.05 was considered statistical significance.

Results: Three significant clinical factors: Epstein-Barr virus-DNA (odds ratio [OR] = 1.748; 95% confidence interval [CI], 0.969-3.171), sex (OR = 2.883; 95% CI, 1.364-6.745), and T stage (OR = 1.853; 95% CI, 1.201-3.052) were identified via univariate and multivariate logistic models. Twenty-four radiomics features were associated with treatment response. Model demonstrated the highest performance among Model , Model , and Model (AUC of training cohort: 0.917 vs. 0.808 vs. 0.795 vs. 0.707; testing cohort: 0.897 vs. 0.755 vs. 0.698 vs. 0.695).

Conclusion: A machine-learning approach using BPNN showed better ability than logistic regression model to predict tumor response to ICT in LANPC.

Evidence Level: 3 TECHNICAL EFFICACY: Stage 2.