Optimizing Esophageal Cancer Radiotherapy Using Delineation with Residual Dynamic Transformer-Nnu-Net: A Comparative Study
Abstract
Purpose
This study proposed the residual dynamic transformer-nnU-Net (RDT-nnU-Net) for automated delineation of esophageal cancer and adjacent organs at risk (OARs) in radiotherapy, aiming to improve accuracy, coverage, and adaptability across radiotherapy strategies.
Methods
A dataset of 83 consecutive esophageal cancer patients undergoing radiotherapy was retrospectively collected. Manual tumor and OAR delineations were performed by radiation oncologists. nnU-Net was identified as the optimal baseline, and enhanced models, including RDT-nnU-Net, were developed. Model performance was evaluated using overlap- and distance-based metrics. Cross-treatment generalizability was assessed with an independent cohort of 328 patients receiving neoadjuvant or definitive radiotherapy for stratified validation.
Results
In the key parameters 2D and 3D Dice similarity coefficient of the clinical target volume, which symbolize tumor delineation accuracy, RDT-nnU-Net outperformed the baseline model and its derivative models (0.905 vs. 0.860–0.905, 0.772 vs. 0.717–0.770, respectively). Additionally, in parameters average surface distance and 95% Hausdorff distance, RDT-nnU-Net also outperformed baseline models and most derivative models (except for nnU-Net + Self-Calibrated Convolution, with 3.37mm vs. 3.73–5.03mm and 26.3mm vs. 29.3–36.5mm, respectively). Moreover, in terms of 3D DSC for OARs, the average DSC values across seven adjacent organs also demonstrated the superiority of RDT-nnU-Net over nnU-Net (0.902 vs. 0.895). During stratified validation, RDT-nnU-Net also demonstrated superior performance compared to baseline models in 3D Dice similarity coefficient during radical radiotherapy (clinical target volume: 0.667 vs. 0.666, gross tumor volume: 0.798 vs. 0.787) and neoadjuvant radiotherapy (clinical target volume: 0.606 vs. 0.605, gross tumor volume: 0.750 vs. 0.741).
Conclusion
RDT-nnU-Net consistently outperformed nnU-Net and other enhanced variants in segmenting esophageal tumors and adjacent OARs, offering improved accuracy, completeness, and generalizability for radiotherapy planning.