Multitask Cyclegan for Joint Image–Dose Learning Improves CBCT Dose Synthesis In Nasopharyngeal Cancer
Abstract
Purpose
This study investigates whether explicitly learning anatomical information in the image domain, when jointly optimized with dose prediction, can improve the accuracy and robustness of CBCT-based dose synthesis.
Methods
This study compares two deep learning strategies for CBCT dose synthesis using 110 nasopharyngeal carcinoma cases with volumetric modulated arc therapy (VMAT) plans (97 training, 13 testing). We implemented: (i) a baseline CycleGAN trained exclusively in the dose domain, and (ii) an innovative multitask CycleGAN that simultaneously learns image translation while optimizing CBCT dose distributions. The multitask architecture features a shared encoder followed by separate decoders for image and dose pathways, thereby enabling joint optimization. The model's core innovation is the introduction of six learnable uncertainty parameters that employ auto-adaptive weighting. Both models were evaluated using image fidelity metrics, and DVH endpoints.
Results
The multitask framework outperformed the baseline CycleGAN across all metrics. Image MAE decreased by 38.6% (0.0063 vs 0.0103), PSNR increased by 2.42 dB, and SSIM improved from 0.91 to 0.94. For GTV D95, mean absolute difference (MAD) decreased from 5.75 to 3.33 Gy, with similar gains for CTV1 (2.79 vs 7.10 Gy) and CTV2 (2.72 vs 8.01 Gy). Parotid V30 MAD decreased from 8.07% to 3.17% (right) and 4.42% to 3.21% (left). Brainstem Dmax MAD improved from 6.10 to 1.70 Gy and spinal cord Dmax from 4.52 to 2.92 Gy. Reduced standard deviations (1.69–3.22 Gy vs 1.88–5.96 Gy) indicated more stable predictions. Comparing dosimetric results against a baseline CycleGAN reveals that joint optimization in image and dose domains yields more accurate and clinically reliable CBCT dose synthesis.
Conclusion
The proposed multitask learning framework jointly aligns image and dose domains, and significantly outperform single domain strategy in synthesizing accurate CBCT dose distributions.