A clinical-radiomics model based on MRI sub-regions of gluteus maximus for recurrence prediction in high-grade serous ovarian cancer.
Authors
Affiliations (8)
Affiliations (8)
- Department of MRI, the First People's Hospital of Yunnan Province, The Affiliated Hospital of Kunming University of Science and Technology, Kunming, China.
- Chuxiong Medical College, Chuxiong, China.
- Department of Obstetrics and Gynecology, The First People's Hospital of Yunnan Province, The Affiliated Hospital of Kunming University of Science and Technology, Kunming, China.
- Department of Radiology, Zhongda Hospital, School of Medicine, Southeast University, Nanjing, China.
- Jiangsu Key Laboratory of Intelligent Medical Image Computing, School of Artificial Intelligence, Nanjing University of Information Science and Technology, Nanjing, China.
- MR Research Collaboration, Siemens Healthineers, Shanghai, China.
- Department of Radiology, The Third Affiliated Hospital of Kunming Medical University, Yunnan Cancer Hospital, Yunnan Cancer Center, Kunming, China.
- Department of MRI, the First People's Hospital of Yunnan Province, The Affiliated Hospital of Kunming University of Science and Technology, Kunming, China. [email protected].
Abstract
To evaluate the predictive value of MRI sub-regional radiomics of the gluteus maximus for recurrence in high-grade serous ovarian cancer (HGSOC) and compare its performance with conventional radiomics and deep learning (DL) models. A multi-center retrospective cohort of 531 patients with HGSOC was analyzed. A U-Net architecture was utilized to automatically segment the region of interest (ROI) of gluteus maximus on T2-weighted imaging (T2WI). Radiomics and DL models were developed using features extracted from the whole ROI. Additionally, sub-regional radiomics models were developed by segregating muscle and fat features through Gaussian Mixture Model (GMM). The optimal predictive model was then integrated with independent clinical predictors to form a combined model. Automatic segmentation accuracy was evaluated using the Dice Similarity Coefficient (DSC). Model performance and clinical utility were assessed using the area under the curve (AUC), net reclassification index (NRI), and integrated discrimination index (IDI). Neoadjuvant chemotherapy (NACT) and poly ADP-ribose polymerase (PARP) inhibitor treatment were identified as independent predictors of HGSOC recurrence (P < 0.05). The U-Net segmentation achieved DSC ranging from 0.726 to 0.903 across cohorts. The sub-regional radiomics model demonstrated superior performance compared to clinical, conventional radiomics, and DL models, achieving AUCs of 0.774 (training), 0.774 (internal validation), 0.717 (external test A), 0.704 (external test B), and 0.717 (external test C). Integrating sub-regional radiomics with clinical factors yielded a combined model with improved AUCs of 0.817, 0.814, 0.746, 0.799, and 0.733, respectively. NRI and IDI analyses confirmed significant net clinical benefits for both the sub-regional radiomics and combined models. The T2WI-based sub-regional radiomics model of the gluteus maximus effectively predicts recurrence in HGSOC patients. The combined model, which integrates sub-regional radiomics features with clinical factors, generally improves predictive accuracy and provides a promising decision-support tool for post-treatment assessment of recurrence risk.