A Radiomics-Based Model to Identify Candidates for De-Intensified Treatment in Patients with Locoregionally Advanced Nasopharyngeal Carcinoma

Abstract

<h3>Purpose/Objective(s)</h3> To develop an optimal radiomics-based model to identify locoregionally advanced nasopharyngeal carcinoma (LA-NPC) patients who can benefit from the de-intensified treatment. <h3>Materials/Methods</h3> LA-NPC patients treated between 2013 and 2018 in Shantou University Medical College Cancer Hospital receiving low dose concurrent cisplatin were divided into training and validation cohort in a 2:1 ratio. For each patient, 107 radiomics features extracted from the primary nasopharyngeal tumor in pre-treatment contrast-enhanced CT scan were analyzed. Clinical model, radiomics model, and combined model were built and validated by Cox proportional hazard analysis. <h3>Results</h3> Training and validation cohort consisted of 66 and 33 patients, respectively. Three predictive independent factors (Flatness representing the sphere-like shape, Mean calculating the average gray level, Gray Level Non-Uniformity (GLDM-GLN) measuring the heterogeneity of gray-level intensity) were remained significantly. The c-index of the clinical model, radiomics model, combined model in the validation cohort was 0.74 (0.64-0.84), 0.73 (0.55-0.91), 0.81 (0.70-0.92), respectively. Combined model, as the optimal model, has a high accuracy with area under curve (AUC) of 0.86 (p=0.01), stratifying patients into low-risk and high-risk group with the 5-year PFS rate 93.5% and 50.8%, respectively (hazard ratio [HR]: 7.37 (1.5-37.3), p=0.006), 5-year OS rate 93.5% and 54.5%, respectively (HR: 7.37 (1.5-37.3), p=0.01). <h3>Conclusion</h3> A radiomics-based combined model can select specific LA-NPC populations for de-intensified treatment without compromising efficacy.