Radiomics feature is a risk factor for locally advanced cervical cancer treated using concurrent chemoradiotherapy based on magnetic resonance imaging: a retrospective study.

Background: Although concurrent chemoradiotherapy (CCRT) is the standard treatment strategy for locally advanced cervical squamous carcinoma (LACSC), there are still individual differences. It is of vital importance to establish a radiomics-based model for predicting overall survival (OS) of LACSC patients treated using CCRT, and evaluating the feasibility of adjuvant chemotherapy (ACT).

Methods: 122 LACSC patients were retrospectively analyzed who underwent pelvic MRI before standard CCRT between January 2013 and September 2016, including 85 patients in training set and 37 patients in testing set. 3D Slicer was used to segment images and extract features. IPMs software was used to select features and construct radscore. We selected the group with the largest area under the curves as the best result from 150 feature subsets and corresponding radscore. A nomogram was established using univariate and multivariate Cox analyses. We used Shapley Additive Explanations (SHAP) for further interpretation of the nomogram. Kaplan-Meier curves demonstrated the associations of radscore and clinical characteristics with OS and ACT.

Results: Radscore was a prognostic factor (P = 0.001) which constructed using 10 radiomics features influencing the OS of patients with LACSC treated using CCRT. The radiomics-clinical model estimated OS (training, C-index: 0.761; testing, C-index: 0.718) more accurately than the clinical (training, C-index: 0.745; testing, C-index: 0.708) and radiomics models (training, C-index: 0.702; testing, C-index: 0.671). Radscore has the greatest impact on the prognosis of LACSC patients. We combined radscore and clinical factors to obtain risk scores. There was a better OS rate among low-risk patients than among high-risk patients (training, P = 0.034; testing, P = 0.003). Compared with CCRT, ACT + CCRT did not improve prognosis (high-risk patients, P = 0.703; all patients, P = 0.425).

Conclusions: Radscore independently predicted OS in LACSC. The radiomics-clinical nomogram improved individualized OS estimation. Patients did not benefit from ACT.