Automated gross tumour volume delineation on T2-weighted magnetic resonance imaging coupled with radiomics for predicting cervical cancer recurrence.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiation Oncology, First Affiliated Hospital of Xi'an Jiaotong University, Xi'an, 710061, China.
- Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, 518048, China.
- Department of Radiation Oncology, First Affiliated Hospital of Xi'an Jiaotong University, Xi'an, 710061, China. [email protected].
- Department of Radiation Oncology, First Affiliated Hospital of Xi'an Jiaotong University, Xi'an, 710061, China. [email protected].
Abstract
Cervical cancer patients are at a high risk of tumour recurrence. Accurate delineation of the primary gross tumour volume (GTV) on pretreatment MR images is crucial for developing personalized prognosis evaluation and treatment. In this study, we collected pretreatment T2-weighted MR data from 306 cervical cancer patients (62 for testing) between 2011 and 2020. We developed a 3D MedNext-7L model for automatic GTV delineation and compared it with the nnUNetV2 model. Manual modifications were made to improve accuracy, and radiomics features were extracted to predict the risk of local recurrence within 5 years. Feature selection was performed using the least absolute shrinkage and selection operator algorithm, and modelling was conducted with the support vector machine algorithm. Additionally, we explored the potential of directly utilizing automatic delineation results for recurrence prediction. The Dice similarity coefficient and sensitivity for the MedNext-7L model in the test set were 0.812 ± 0.144 and 0.848 ± 0.128, respectively, surpassing those of nnUNetV2 at 0.803 ± 0.154 and 0.823 ± 0.155, respectively. Using manually refined GTVs, our model achieved AUC values of 0.875 and 0.811 in the fivefold cross-validation set and test set, respectively. On the basis of the automatically delineated GTVs, our model achieved AUC values of 0.786 and 0.730 in the fivefold cross-validation set and test set, respectively. The 3D MedNext-7L model outperforms nnUNetV2 in GTV delineation on pretreatment T2-weighted MR images. Rapid manual modifications combined with radiomics and machine learning techniques can effectively predict recurrence risk. The automatic delineation results show promise for direct application in recurrence prediction, indicating a feasible pathway towards a fully automated predictive system.