A Cgan-Based Spatially-Resolved Prediction of Lung Tumor Response to Chemoradiation
Abstract
Purpose
Lung cancer is a leading global malignancy with high mortality. Radiotherapy is a critical treatment; however, current planning often suffers from subjective dose settings and side effects. This study aims to use a Conditional Generative Adversarial Network (cGAN) to accurately predict spatially-resolved response, assisting doctors in optimizing doses and promoting precise adaptive radiotherapy.
Methods
For 25 patients with locally advanced NSCLC (NCT02773238), PET/CT images and dose distributions were retrospectively analyzed. Masked tumor volumes were z-sliced, and PreSUV was channel-merged with radiation dose to form dual-channel Pre-Dose inputs. A cGAN framework was constructed, consisting of a U-Net generator with multi-level residual blocks to synthesize MidSUV maps, and a PatchGAN discriminator to distinguish real from generated images. The model was optimized using a hybrid objective combining voxel-wise Mean Squared Error (MSE) and adversarial loss. A rigorous leave-one-out cross-validation protocol was implemented.t Performance was evaluated using voxel-wise Root Mean Square Error (RMSE), Spearman correlation across patients stratified into high, medium, and low response groups, with component analyses validating the architectural design.
Results
The proposed framework achieved a mean voxel-wise SUV RMSE of 1.34 and Spearman correlation of 0.86 for the entire cohort. Stratified analysis showed the high-response group yielded best response ratio prediction (Ratio RMSE:0.17), while the low-response group exhibited superior absolute MidSUV fidelity (RMSE:1.20). Component analysis confirmed the design: replacing residual blocks with standard layers increased the RMSE to 1.38, while removing the adversarial component caused further degradation (RMSE increased to 1.51, correlation dropped to 0.74), confirming the necessity of the proposed residual backbone and adversarial training.
Conclusion
The proposed framework robustly predicts spatially-resolved tumor response. Integrating adversarial training and residual connections effectively improves accuracy compared to standard generators. By capturing response variations across patient groups, this tool supports dose optimization and facilities precise adaptive radiotherapy.