ESTRO 2024 - Abstract Book

S4028

Physics - Inter-fraction motion management and offline adaptive radiotherapy

ESTRO 2024

1296

Poster Discussion

Onboard synthetic 4D-MRI generation for target localization on a conventional X-ray guided LINAC

Paulo Quintero, Wendy Harris, Laura Cervino

Memorial Sloan Kettering, Medical Physics, New York, USA

Purpose/Objective:

Accurate target localization is crucial for abdominal cancers due to respiratory motion, which can cause treatment errors. Magnetic Resonance Imaging (MRI)-guided radiation therapy (MRgRT) is superior for these treatments due to superior soft tissue contrast in MRI compared to Cone-beam Computed Tomography (CBCT) from a conventional LINAC. The development of on-board synthetic MRI using a conventional LINAC would substantially improve target localization accuracy for abdominal radiotherapy. Previous AI-based solutions to generate synthetic MRI images for radiotherapy applications are oriented to organ segmentation and based on U-net or generative-adversarial network (GAN) architectures, requiring big training datasets, expensive computational resources, and, most importantly, unclear model’s reliability evaluation [1]. Furthermore, as the few reported synthetic MRI methods with potential adaptive radiotherapy applications have mentioned [2-5], the ground truth MRI is not possible to achieve due to the technological limitations of having CBCT and MRI in the same on-board conditions within the treatment room. Thus, using 4D virtual phantom sets, we proposed a novel on-board synthetic 4D-MRI generation method to be applied to an X-ray guided LINAC that might be efficiently adapted to a radiotherapy workflow to improve abdominal adaptive radiotherapy treatments when MR-guided LINAC is not available. Synthetic MRI generation : Using the virtual phantom 4D-XCAT [6], a 4D-MRI and its corresponding 4D-CT for treatment planning were generated. Deformable image registration (DIR) was performed between the end of expiration (EOE) phase of the MRI and the other phases to obtain the 4D deformation field map (4D-DFM). The deformation eigenvalues and eigenvectors were calculated implementing Principal component analysis (PCA) applied to 4D-DFM. Then, after sampling 1000 times the eigenvalues, 1000 new deformations were applied to the CT’s EOE phase, creating the training dataset. Finally, a CNN model was trained to predict the eigenvalues from a 4D-CBCT, that will be applied to deform the original EOE-MRI, creating the synthetic 4D-MRI. Figure 1 shows the overall workflow of the methods. Additionally, the model robustness was tested changing specific respiratory features in new different 4D-CBCT scenarios: respiratory cycle (RP, [s]), diaphragm amplitude (DA, [mm]), and chest wall amplitude (CW, [mm]). Model evaluation : The CNN model performance was evaluated with RMSE and MSE, the synthetic image quality was evaluated with SSIM and nRMSE, and the liver and target volume structures were evaluated with volumetric dice coefficient (DC), volume percent difference (VPD), and center-of-mass shift (COMS). Material/Methods:

Results: