Improving outcome prediction in oropharyngeal carcinoma through the integration of diffusion-weighted magnetic resonance imaging radiomics

Abstract

Background and purpose

Locoregional recurrence (LRR) is the primary pattern of failure in head and neck cancer (HNC) following radiation treatment (RT). Predicting an individual patient’s LRR risk is crucial for pre-treatment risk stratification and treatment adaptation during RT. This study aimed to evaluate the feasibility of integrating pre-treatment and mid-treatment diffusion-weighted (DW)-MRI radiomic parameters into multivariable prognostic models for HNC.

Materials and methods

A total of 178 oropharyngeal cancer (OPC) patients undergoing (chemo)radiotherapy (CRT) were analyzed on DW-MRI scans. 105 radiomic features were extracted from ADC maps. Cox regression models incorporating clinical and radiomic parameters were developed for pre-treatment and mid-treatment phases. The models’ discriminative ability was assessed with the Harrel C-index after 5-fold cross-validation.

Results

Gray Level Co-occurrence Matrix (GLCM)-correlation emerged as a significant pre-treatment radiomic predictor of locoregional control (LRC) with a C-index (95 % CI) of 0.66 (0.57–0.75). Significant clinical predictors included HPV status, stage, and alcohol use, yielding a C-index of 0.70 (0.62–0.78). Combining clinical and radiomic data resulted in a C-index of 0.72 (0.65–0.80), with GLCM-correlation, disease stage and alcohol use as significant predictors. The mid-treatment model, which included delta (Δ) mean ADC, stage, and additional chemotherapy, achieved a C-index of 0.74 (0.65–0.82). Internal cross-validation yielded C-indices of 0.60 (0.51–0.69), 0.56 (0.44–0.66), and 0.63 (0.54–0.73) for the clinical, combined, and mid-treatment models, respectively.

Conclusion

The addition of Δ ADC improves the clinical model, highlighting the potential complementary value of radiomic features in prognostic modeling.