MRI-based deep learning models for automatic segmentation and pretreatment risk stratification in operable cervical squamous cell carcinoma: a multicenter study

Background Precisely estimating risk stratification prior to initial treatment is crucial to guide clinical treatment decision-making in patients with operable cervical cancer. This study aimed to develop and validate deep-learning (DL) models using multi-parameter magnetic resonance imaging (MRI) for automatic segmentation and risk status prediction in operable cervical squamous cell carcinoma (CSCC). Methods Between June 2015 and May 2024, a total of 408 patients were enrolled from four hospitals. The nnUNetV2 architecture was designed for automatic tumor segmentation. For classification task, the 3D DL and 2.5D DL model, clinical model were developed. Two fusion models were developed: the feature-based Combined model (FeatureMerge) and the probability-based Combined model (ProbML). The segmentation performance was assessed using the Dice similarity coefficient (DSC), and the model’s diagnostic performance was assessed using area under the receiver operating characteristic curve (AUC). Results The segmentation model performed well in the internal testing cohort and external testing cohort, with average DSCs of 0.829 and 0.762, respectively. For the prediction of risk stratification, the 2.5D DL model had superior discriminative ability compared to 3D DL model and clinical model. Combined model (ProbML) exhibited improved predictive capacity in the internal testing cohort (AUC, 0.865; 95% CI, 0.773–0.958) and external testing cohort (AUC, 0.805; 95% CI, 0.713–0.897). Conclusion The proposed Combined model (ProbML), which incorporates clinical characteristics, 3D DL and 2.5D DL features, can be used to predict risk stratification in operable CSCC. In addition, the nnUNetV2-based deep-learning model can accurately segment cervical cancer.