Author: Derek Tang, Susu Yan 👨🔬
Affiliation: Massachusetts General Hospital 🌍
Purpose: To evaluate the performance of a multi-task automated-segmentation and synthetic CT generation model (sCT) and investigate its application in an adaptive proton therapy workflow.
Methods: A cycle-consistent generative adversarial network (CycleGAN) is designed and trained to generate sCT images and produce multi-channel segmentations simultaneously from cone-beam CT (CBCT) acquisitions using a modified GAN and cycle-consistent softmax cross-entropy segmentation loss function. CBCT and CT images with physician contours from 186 patients treated for prostate cancer on an Elekta Agility Linear Accelerator were used in training, validation, and testing of the model. An additional set of images from 20 patients that received adaptive treatments on a Varian ETHOS system are used in testing the model’s segmentation performance.
The mean absolute error (MAE), root-mean-squared error (RMSE), structural similarity index metric (SSIM), and DICE scores were used to determine the image quality of sCT and segmentation accuracy of outputs relative to a deformably registered ground truth. These results were compared to other proposed multi-task frameworks and a sequential workflow deploying TotalSegmentator. sCT images were evaluated dosimetrically for proton radiotherapy using RayStation.
Results: The MAE, RMSE, SSIM, and DICE scores of the outputs from the proposed model are 17.04 ± 4.66 HU, 56.67 ± 7.41 HU, 0.74 ± 0.05, and 0.85 ± 0.04 respectively, and outperform the outputs of the alternative multi-task frameworks and sequential workflow. The passing rate for adaptive plans using multi-task sCT outputs with a gamma criterion of 2mm/2% is 95.89 ± 3.95%.
Conclusion: The proposed multi-task network helps optimize adaptive radiotherapy workflows by successfully implementing simultaneous segmentation and unsupervised image translation. The cycle-consistent training improves upon established multi-task architectures in literature by also incorporating segmentation specific losses. This model aims to reduce the amount of time a patient is on the treatment table and minimize uncertainties in adaptive proton treatments.