Radiomics-Guided Automatic Delineation for Clinical Target Volume of Endometrial Cancer: Limited Sample Multicenter...

The inconsistent delineation of the clinical target volume (CTV) in postoperative pelvic radiation therapy for endometrial carcinoma across centers poses a challenge for deep learning-based segmentation due to differing definitions of the internal target volume. This study aimed to develop an effective method to address multi-institutional variations in CTV delineation, even under the constraints of limited data availability.

A total of 207 simulated computed tomography cases of patients with endometrial cancer across 5 centers were retrospectively collected. Within each center, the data were divided into support, query, and test sets. Each center was sequentially designated as the target center for fine-tuning and testing, while the remaining 4 centers were used for model training to validate the superiority of the proposed method. In addition, 26 cases from an external center were used exclusively for fine-tuning and testing. Radiomics features were extracted to analyze differences in CTV delineation and images across centers. A random forest classifier was trained to identify the most important radiomics features. Using these features as guidance, a model-agnostic meta-learning (MAML) strategy was applied to pretrain a 3-dimensional U-Net radiomics-guided MAML (MAML-r) model, which was subsequently fine-tuned on each target center's data. The performance of the proposed MAML-r method was compared with direct 3-dimensional U-Net training and transfer learning models. Evaluation metrics included the Dice similarity coefficient (DSC), the 95th percentile Hausdorff distance (HD95), and the average symmetric surface distance (ASSD), supplemented by qualitative assessments from clinical experts using a 4-point scoring system.

Eight important features were identified from a total of 107 radiomics features, which showed significant differences across centers (P < .01). The MAML-r model yielded meaningful results, achieving a mean ± SD DSC of 0.818 ± 0.058, a mean ± SD HD95 of 9.314 ± 3.648 mm, and a mean ± SD ASSD of 2.772 ± 1.090 mm. The model also earned an average blinded expert evaluation score of 3.24, significantly outperforming all other models. Notably, improved performance was observed in the external test cohort, with corresponding mean ± SD values for DSC, HD95, and ASSD of 0.886 ± 0.012, 5.203 ± 1.435 mm, and 1.512 ± 0.334 mm, respectively. Furthermore, the MAML-r model achieved the shortest mean ± SD CTV modification time of 3.8 ± 1.2 minutes. Given the variations in CTV contouring styles across centers and the limited sample size, the MAML-r model demonstrated superior performance and adaptability compared with the other models.

This study introduces a novel MAML-r framework for few-shot, multicentric CTV segmentation tasks in postoperative pelvic radiation therapy for endometrial carcinoma, significantly mitigating performance degradation caused by interinstitutional variations in delineation styles and data scarcity. The proposed approach offers a promising solution to these persistent clinical challenges.