Uncertainty-Aware Simultaneous Synthesis and Segmentation Via Evidential Deep Learning

Daily anatomical changes during radiotherapy demand accurate imaging and segmentation to support adaptive treatment. We propose Simultaneous Evidential Network for Synthesis and SEgmentation (SENSSE), a deep learning framework that jointly performs synthetic CT generation and anatomical segmentation from daily CBCT or MR images, while also estimating voxel-wise uncertainty. The architecture integrates a shared encoder and task-specific decoders, trained with a combined loss that promotes cross-task supervision, allowing segmentation to guide synthesis with anatomical priors and synthesis to complement segmentation with structural fidelity. To improve interpretability, the framework incorporates evidential modeling to estimate both data and model-related uncertainty in a single forward pass, resulting in spatial maps that highlight regions of low confidence. These maps support quality assurance and can inform manual review or adaptive thresholds. Evaluations on head and neck and pelvic datasets showed that SENSSE achieved competitive synthesis performance, improved OAR segmentation accuracy and better calibrated uncertainty estimates compared with deterministic and sampling-based baselines. The multitask design improves anatomical consistency and reduces artifacts, while the uncertainty-aware outputs provide additional information to support clinical decision-making. Overall, the framework offers a robust and efficient solution for adaptive radiotherapy workflows, combining multitask learning with uncertainty quantification to support safer and more informed treatment adaptation.