Development and validation of a magnetic resonance imaging-based model for the prediction of progression-free survival before initial treatment of hypopharyngeal cancer treated with chemoradiation

Objective To develop a radiomics-clinical model for prediction of the progression-free survival (PFS) in hypopharyngeal squamous cell carcinoma (HPSCC) patients treated with concurrent chemoradiotherapy (CCRT). Materials and methods One hundred and thirteen patients with hypopharyngeal cancers treated by CCRT were enrolled and randomly divided into training (N = 80) and validation (N = 33) groups. Python’s PyRadiomics toolkit was used to extract radiomic features from magnetic resonance diffusion-weighted imaging (DWI) sequences, which included four categories: shape (14), first-order (230), texture (306), and wavelet (738). All feature preprocessing steps—including normalization, intraclass correlation coefficient (ICC) filtering, Pearson correlation filtering, and feature selection—were performed exclusively within the training set, with the same preprocessing parameters subsequently applied to the validation set. The least absolute shrinkage and selection operator (LASSO) penalized Cox regression algorithm with ten-fold cross-validation was used to screen key radiomic features and construct the Rad-score. We used Kaplan-Meier analysis and log-rank tests to assess the ability of the radiomic signature to stratify PFS risk, and constructed a radiomics nomogram to predict individual PFS probabilities at 1 and 2 years. Results The radiomics model consisting of six important features performed well in classifying patients into high- and low-risk groups of progression in both the training and test groups (log-rank test, p = 0.006 in the training cohort; p = 0.023 in the validation cohort, using the cutoff value of 0.025 determined from the training set). Age and T stage were chosen as clinically important variables based on univariate Cox regression analysis (p < 0.05). The radiomics nomogram demonstrated good differentiation ability, with C-indices of 0.904 (95% confidence interval [CI], 0.816–0.992) and 0.858 (95% CI, 0.690–1.000) in the training and validation cohorts, respectively. The time-dependent AUCs for predicting 1-year and 2-year PFS were 0.867 and 0.837, respectively, in the training cohort, and 0.737 and 0.715 in the validation cohort. Conclusion The multiparametric DWI-based radiomics model may serve as a potential tool to predict PFS in patients with HPSCC treated with CCRT after further prospective validation.