2204 - Radiomics-Guided Deep Learning-Based Framework for Few-Shot Automatic CTV Delineation of Endometrial Carcinoma in Multi-Center
Presenter(s)
A. Qu1, X. Zhang1, H. Yang2, W. Xiong3, L. Wei4, Y. Zhuo5, X. Xue6, Y. Song7, L. Jia8, P. Jiang1, and J. Wang1; 1Department of Radiation Oncology, Peking University Third Hospital, Beijing, China, 2United Imaging Research Institute of Intelligent Imaging, Beijing, China, 3Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China, 4Department of Radiation Oncology, First Affiliated Hospital of Air Force Medical University, Xi'an, China, 5Zhangzhou Hospital of Fujian Province, zhangzhou, China, 6the Second Hospital Hebei Medical University, shijiazhuang, China, 7Yantai Yuhuangding Hospital, yantai, China, 8Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China
Purpose/Objective(s): The CTV of postoperative radiotherapy for endometrial carcinoma(EC) includes the vaginal stump, proximal vagina, and paravaginal/parametrial regions, lymphatic drainage areas. Different parts are affected differently by organ motion. In clinical practice, they are usually contoured as a whole, which results in the actual CTV encompassing the internal target volume (ITV). These results in variations in delineation practices among different centers and individual radiation oncologists, despite their adherence to the same clinical guidelines. Such inconsistencies in delineation can significantly impede the performance of deep learning-based segmentation algorithms, particularly when faced with limited training datasets. Our purpose is to find a method to effectively accommodate multi-institutional variations, even under the constraints of scarce data availability.
Materials/Methods: Contrast-enhanced CT images and CTVs from 187 postoperative endometrial cancer patients from five centers were collected. Radiomics features were extracted to analyze differences in CTV delineation and image among the centers. A Random Forest Classifier (RFC) was trained to identify the most important radiomics features. A Model-Agnostic Meta-Learning (MAML) strategy was employed to pre-train a 3D-UNet model using radiomics features as guidance (MAML-r). The model was fine-tuned on each target center and evaluated using five-fold cross-validation, where each center was sequentially treated as the target center and the others as meta-centers. The performance of MAML-r was assessed using Dice similarity coefficient (Dice), 95th percentile Hausdorff Distance (HD95), and Average Symmetric Surface Distance (ASSD).
Results: The top 8 important features were identified from 107 radiomics features, which showed significant differences across centers (p < 0.01). The MAML-r method outperformed direct 3D-UNet training and transfer learning methods, with a mean Dice of 0.818±0.058, a mean HD95 of 9.314±3.648, and a mean ASSD of 2.772±1.090.
Conclusion: The study introduces a radiomics-guided MAML method for multi-center CTV segmentation in postoperative pelvic radiotherapy of EC. This approach successfully addresses the challenges of limited data availability and variability in CTV delineation across different centers. The MAML-r method provides a framework for improving segmentation accuracy and consistency in clinical practice, even with scarce and heterogeneous datasets.