Integration of radiomics, habitat imaging, and deep learning for MRI-based prediction of parametrial invasion in cervical cancer: A dual-center study.
Authors
Affiliations (3)
Affiliations (3)
- Dalian Medical University, Dalian, Liaoning 116044, China; Department of Gynecology and Obstetrics, Dalian Municipal Central Hospital, Affiliated of Dalian University of Technology, Dalian 116033, China.
- Department of Gynecology and Obstetrics, Second Affiliated Hospital of Dalian Medical University, Dalian 116000, China.
- Department of Gynecology and Obstetrics, Second Affiliated Hospital of Dalian Medical University, Dalian 116000, China. Electronic address: [email protected].
Abstract
To assess the diagnostic performance of radiomics, habitat imaging, and 2.5D deep learning models for MRI-based prediction of parametrial invasion in cervical cancer, and to evaluate the clinical utility of a multimodal integrated model. This dual-center retrospective study included 290 patients with FIGO stage IB1-IIB cervical cancer who underwent preoperative MRI. Patients from Center A (n = 227) were divided into training and validation cohorts, while patients from Center B (n = 63) comprised the external test cohort. Radiomic features were extracted, habitat imaging was performed using k-means clustering, and a 2.5D deep learning model incorporated adjacent slices. Feature selection was conducted using Pearson correlation and LASSO regression. Machine learning models were developed, and an integrated model was constructed. Model performance was evaluated using AUC and accuracy. AUCs were compared with DeLong tests, calibration was assessed with the Hosmer-Lemeshow test, and clinical utility was evaluated with decision curve analysis. The integrated model outperformed all individual models, achieving AUCs of 0.973, 0.901, and 0.906 in the training, validation, and external test cohorts, respectively. Among individual models, the deep-learning model showed the highest AUCs (0.954, 0.803, 0.833), followed by habitat imaging (0.860, 0.811, 0.843). In the external test cohort, the peritumoral radiomics model outperformed the intratumoral model (0.843 vs. 0.719). The clinical model showed the lowest performance. Hosmer-Lemeshow tests indicated good calibration, and decision curve analysis confirmed superior clinical utility of the integrated model. The multimodal integrated model, combining radiomics, habitat imaging, 2.5D deep learning, and clinical features, demonstrated superior predictive performance for parametrial invasion in cervical cancer compared with individual models. This approach may enhance preoperative assessment, guide clinical decision-making, and optimize treatment strategies.