Generative evidential synthesis with integrated segmentation framework for MR-only radiation therapy treatment planning.

Background: Radiation therapy (RT) planning is a time-consuming process involving the contouring of target volumes and organs at risk, followed by treatment plan optimization. CT is typically used as the primary planning image modality as it provides electron density information needed for dose calculation. MRI is widely used for contouring after registration to CT due to its high soft tissue contrast. However, there exists uncertainties in registration, which propagate throughout treatment planning as contouring errors, and lead to dose inaccuracies. MR-only RT planning has been proposed as a solution to eliminate the need for CT scan and image registration, by synthesizing CT from MRI. A challenge in deploying MR-only planning in clinic is the lack of a method to estimate the reliability of a synthetic CT in the absence of ground truth. While methods have used sampling-based approaches to estimate model uncertainty over multiple inferences, such methods suffer from long run time and are therefore inconvenient for clinical use.

Purpose: To develop a fast and robust method for the joint synthesis of CT from MRI, estimation of model uncertainty related to the synthesis accuracy, and segmentation of organs at risk (OARs), in a single model inference.

Methods: In this work, deep evidential regression is applied to MR-only brain RT planning. The proposed framework uses a multi-task vision transformer combining a single joint nested encoder with two distinct convolutional decoder paths for synthesis and segmentation separately. An evidential layer was added at the end of the synthesis decoder to jointly estimate model uncertainty in a single inference. The framework was trained and tested on a dataset of 119 (80 for training, 9 for validation, and 30 for test) paired T1-weighted MRI and CT scans with OARs contours.

Results: The proposed method achieved mean ± SD SSIM of 0.820 ± 0.039, MAE of 47.4 ± 8.49 HU, and PSNR of 23.4 ± 1.13 for the synthesis task and dice similarity coefficient of 0.799 ± 0.132 (lenses), 0.945 ± 0.020 (eyes), 0.834 ± 0.059 (optic nerves), 0.679 ± 0.148 (chiasm), 0.947 ± 0.014 (temporal lobes), 0.849 ± 0.027 (hippocampus), 0.953 ± 0.024 (brainstem), 0.752 ± 0.228 (cochleae) for segmentation-in a total run time of 6.71 ± 0.25 s. Additionally, experiments on challenging test cases revealed that the proposed evidential uncertainty estimation highlighted the same uncertain regions as Monte Carlo-based epistemic uncertainty, thus highlighting the reliability of the proposed method.

Conclusion: A framework leveraging deep evidential regression to jointly synthesize CT from MRI, predict the related synthesis uncertainty, and segment OARs in a single model inference was developed. The proposed approach has the potential to streamline the planning process and provide clinicians with a measure of the reliability of a synthetic CT in the absence of ground truth.