Predicting Response to Radiotherapy in Locally Advanced Cervical Squamous Cell Carcinoma Based on Multisequence MRI and Clinical Variables.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiation Oncology, Zhangzhou Hospital Affiliated to Fujian Medical University, Zhangzhou, Fujian 350600, China (Y.Z., Y.W., Y.L., L.Q., D.K., Y.X.).
- Department of Radiology, Zhangzhou Hospital Affiliated to Fujian Medical University, Zhangzhou, Fujian, China (F.L.).
- School of Informatics, Xiamen University, Xiamen, Fujian, China (R.Z.).
- Department of Pathology, Zhangzhou Hospital Affiliated to Fujian Medical University, Zhangzhou, Fujian, China (C.Z.).
- Department of Radiation Oncology, Zhangzhou Hospital Affiliated to Fujian Medical University, Zhangzhou, Fujian 350600, China (Y.Z., Y.W., Y.L., L.Q., D.K., Y.X.). Electronic address: [email protected].
Abstract
Locally advanced cervical cancer (LACC) remains a major cause of cancer-related mortality among women. We developed and tested a machine learning model integrating radiomic and clinical features to predict tumor shrinkage rate (TSR) after external beam radiotherapy (EBRT). This study retrospectively enrolled 248 patients with histopathologically confirmed locally advanced cervical squamous cell carcinoma from April 2020 to April 2025. The cohort was split randomly into a training set (n = 173) and a validation set (n = 75) for internal testing. A deep learning model extracted deep features from T1-weighted, T2-weighted, T2-SPAIR, and diffusion-weighted imaging (DWI) pretreatment Magnetic resonance imaging (MRI) sequences. Fusion combined these radiomic features with clinical variables. Conditional weighted gradient boosting survival (CWGBS), eXtreme Gradient Boosting (XGBoost), gradient boosting survival tree (GBST), and random survival forest (RSF) models were trained on the fused dataset. Model performance metrics comprised area under the receiver operating characteristic curve Area Under the Curve (AUC), accuracy, sensitivity, and specificity. The machine-learning radiomics-clinical fusion models demonstrated robust performance in predicting TSR after EBRT. The AUCs of the four fusion models ranged from 0.824 to 0.863, outperforming models based solely on DWI (AUC 0.711-0.754), T1WI (AUC 0.681-0.737), T2WI (AUC 0.694-0.825), or T2-SPAIR (AUC 0.714-0.827). Among them, the RSF model achieved the highest accuracy (0.920), sensitivity (0.937), specificity (0.833), and AUC (0.863). A deep-learning radiomics-clinical model demonstrated robust capability for predicting TSR after EBRT in LACC. Early identification of high-risk patients supports individualized treatment planning.