Feasibility study of three-dimensional dose prediction of esophageal cancer radiotherapy based on CGAN
Qichao Hu1, Yanrui Li1, Youyi Tao1
1Department of Radiation Oncology, Donghai County People's Hospital affiliated to Kangda College of Nanjing Medical University, Lianyungang, Jiangsu, China.
Objective:
To evaluate the feasibility of a Conditional Generative Adversarial Network (CGAN)-based model (2UDCNN-AT-GAN) for predicting 3D radiotherapy dose distributions in esophageal cancer, and to assess its potential utility as a guidance tool for Intensity-Modulated Radiotherapy (IMRT) plan design.
Methods:
A CGAN-based model was trained on 95 esophageal cancer IMRT plans and tested on 15 plans. The model integrated a dual-path encoder-decoder, attention-gated fusion, and Dose-volume histogram (DVH)-based conditioning. Predicted dose distributions were compared with manual clinical plans using DVH metrics, plan evaluation parameters, and error analysis.
Results:
All predicted plans met clinical constraints (100% qualification rate). Predicted and manual plans identified concordant near-limit organs at risk (OARs) cases. This consistency mirrored the critical OAR constraints that govern target dose distribution. The mean prediction error was 2.88 ± 2.88% with target errors of 1.00 ± 1.00%. The homogeneity index (0.08 ± 0.02) matched that of manual plans, while the conformity index (CI) was slightly lower (p = 0.0034). The mean Dice Similarity Coefficient (DSC) was 0.916.
Conclusion:
In this feasibility study, the 2UDCNN-AT-GAN model generated preliminary clinically acceptable dose predictions that met basic clinical constraints and showed spatial similarity to manual plans. A systematic ablation study confirmed that the integrated dual-path and attention-gated architecture outperforms individual components, establishing proof-of-concept feasibility for esophageal IMRT planning guidance. However, systematic target underdosage and the single-institution scope limit its current role to a preliminary planning reference.
