Prior-guided automatic delineation of post-radiotherapy gross tumor volume for esophageal cancer

Integrating post-radiotherapy (RT) CT into longitudinal esophageal cancer response models substantially improves predictive accuracy. However, manual delineation of gross tumor volume (GTV) on post-RT CT is both labor-intensive and time-consuming.

We propose a novel deep learning-based framework that integrates medical physics priors-pre-RT GTV contours and radiotherapy dose distributions-to automatically delineate post-RT GTV.

A multicenter retrospective cohort of 294 EC patients (225 training, 45 internal validation, 24 external validation) was assembled. Pre-RT CT scans, GTV contours, and dose map were co-registered and cropped to 256 × 256. We implemented an nnU-Net v2 backbone, incorporating high dose region and pre-RT GTV priors via element-wise multiplication and element-wise addition to guide feature extraction. Performance was evaluated using anatomical (Dice, IoU, HD95, ASSD, Precision, Recall) and radiomics analyses (ICC, Pearson correlation, LASSO-Cox, C-index) across internal and external cohorts.

In cross-validation, the optimal fold achieved DSC = 0.7809 ± 0.1310, IoU = 0.6486 ± 0.1507, HD95 = 3.6321 ± 2.0942, and ASSD = 1.9673 ± 1.0352 (p < 0.0167 vs. state-of-the-art models). Ablation studies demonstrated that combining two types of medical physics priors outperformed single-prior or no-prior models (internal: DSC = 0.7723 ± 0.1290; external: DSC = 0.7545 ± 0.1058). Radiomic features extracted from automated contours exhibited high reproducibility (78.6% with ICC > 0.75) and strong concordance with manual features (R >  0.8), yielding comparable prognostic performance (C-index Δ nonsignificant).

By embedding medical physics priors into a self-configuring nnU-Net v2, our method achieves accurate and robust automated delineation of post- RT GTV in EC across multiple centers. This approach has potential to facilitate the construction of treatment response prediction models.