ESTRO 2025 - Abstract Book

S831

Clinical - Gynaecology

ESTRO 2025

Material/Methods: In this retrospective single-center study, we collected pre-treatment T2-weighted MR images of 306 patients (62 in the test set) of cervical cancer between November 2011 and May 2020 with their outcome information. The GTV on MR images was manually modified based on the masks generated by the automatic delineation model. Patients were labeled as positive if they experienced local recurrence or persistence within five years after treatment. A seven-layer 3D MedNext-s model was developed for the automatic delineation of GTV, and its performance was compared with nnUNetV2. Radiomic features were extracted from both the original and wavelet-filtered MR images of the GTV using PyRadiomics. Feature selection was performed using the Least Absolute Shrinkage and Selection Operator algorithm, followed by modeling with the Support Vector Machine algorithm to accomplish recurrence prediction. The performance of the automatic delineation model was evaluated using metrics including Dice Similarity Coefficient (DSC), sensitivity, and Average Symmetric Surface Distance (ASSD). The performance of recurrence prediction model was assessed in terms of accuracy, specificity, sensitivity, and the area under the receiver operating characteristic curve (AUC). Furthermore, we investigated the potential of employing unaltered automatic delineation outcomes for both training and test sets in recurrence prediction.

Results:

For the automatic delineation, the results are summarized in Table 1. The DSC on the test set for the MedNext-s model was 0.812±0.144, which is superior to that of nnUNetV2 at 0.803±0.154. For recurrence prediction, the results are demonstrated in Table 2. Using manually modified GTVs, our model achieved AUCs of 0.875 and 0.811 on the five-fold cross-validation set and test set, respectively. Furthermore, based on automatically delineated GTVs, our model achieved AUCs of 0.786 and 0.730 on the five-fold cross-validation set and test set, respectively. Conclusion: The 3D MedNext-s model demonstrates an exceptional ability in GTV delineation on pre-treatment T2-weighted MR images of cervical cancer patients. Following rapid manual modifications based on the automatic delineation results, combined with radiomics and machine learning techniques, it can be utilized for recurrence prediction. Moreover, the automatic delineation results demonstrate promising potential for direct application in recurrence prediction, suggesting a feasible pathway towards a fully automated system for local recurrence prediction.

Keywords: cervical cancer, magnetic resonance, recurrence