ESTRO 2021 Abstract Book

S1406

ESTRO 2021

Purpose or Objective Magnetic resonance (MR) imaging is a growing modality in terms of use in radiation therapy. Used mainly in the past as secondary images for brain tumor and organs at risk annotation as for optic chiasm or hippocampus, there is now a real interest to use it as primary image for dose planning. However, in order to determine the optimal treatment plan, information related to the tissue properties are still necessary. Therefore, Computed Tomography (CT)-equivalent representations are needed for dose calculations. To this end, we propose a novel self-supervised generative adversarial deep learning approach to generate synthetic-CTs from MR images that can learn from unaligned MR-CT pairs. Our approach can generate synthetic CTs from T1 weighted brain MRIs (including gadolinium enhanced scans) in real-time and thus seamlessly integrates to the MR-Linear accelerator workflow. Materials and Methods A dataset of 1242 T1 weighted brain MRI scans and the corresponding CT scans from multiple devices and manufacturers was collected from various sites. We deployed a two phase learning pipeline involving three key steps: (i) Conditional generative adversarial deep learning based weakly supervised cross modality image synthesis to generate synthetic CT priors from MR images (ii) Highly accurate alignment of CT to the MRI using weak priors via mono-modal multi-metric deformable registration with a combination of intensity driven and intensity agnostic metrics to generate paired data (iii) Synthetic CT generation with the self-paired data using deep generative adversarial networks. Multiple networks were trained using different whole body scans as reference space. Each of them relied on a different random separation between training (80%) and validation (20%) subsets. Evaluation was performed on a held out test set, 25% the size of the training and validation data. Mean absolute error (MAE) on the patient head was computed for performance evaluation. Results We report a test MAE of 77.22 ± 17.19 HU on the patient body. For 39 cases from the test set, MAE was specifically computed for bone, air and water regions, giving 31.57±2.37 for water, 189.87±19.58 for bones and 206.56±16.08 for air. We are also planning a dosimetric impact study on these 39 patients. For qualitative assessment an example of the input MRI, the corresponding CT and the generated synthetic CT is shown in the figure. Conclusion We introduced an innovative ensembling AI-driven strategy for synthetic CT generation that fights the use of multi modal registration with a cascading approach. It helps to generate synthetic CTs that are suitable for clinical use thanks to its intensity and structure preserving learning strategy that can generate high quality, sharp synthetic CTs with accurate hounsfield values. Moreover, our approach inherits robustness and good generalization properties through an ensembling principle done on anatomically consistent sub-spaces. PO-1681 Effectiveness of a Cranial Distortion Correction Software Using A Novel Measurement Method T. Belloeil-Marrane 1 , A. Gutierrez 1 , J. Smeulders 1 , T. Gevaert 1 , M. De Ridder 1 1 UZ Brussel, Radiotherapy, Brussels, Belgium Purpose or Objective The accuracy of a stereotactic treatment is primarily limited by the least accurate process in the whole chain of events. QA is often performed on the dose delivery and planning section rather than the localization. MRI datasets are subjected to distortions, due to the nonlinearity of gradient fields, and may cause incorrect target definition. This study aimed to analyze the impact of a patient-specific algorithm, Cranial Distortion Correction Elements (Brainlab, München, Germany), rather than a manufacturer-specific, to correct spatial distortion in cranial magnetic resonance images by using a novel software-only evaluation paradigm independent from potential bias linked to the image modalities and with defined increment values of the distortion and noise present in clinical images. Materials and Methods A non-bias simulated T1 MRI normal brain dataset (Brainweb) is utilised to create synthetic CT. By introducing controlled distortion in simulated datasets, we can evaluate the influence of the Noise (Gaussian noise percent multiplied by the brightest tissue intensity) and Intensity non-uniformity ("RF"). These image sets are referenced to the origin of the synthetic CT in order to eliminate geometrical shifts. For this study, we have used 17 MRI datasets ranging from 0 to 9% noise and from 0 to 40% RF. These MRIs were corrected using the synthetic CT as a base modality for the distortion correction and equal number of corrected MRIs were generated. To evaluate the impact of the distortion correction, each image set, non-corrected and corrected, was compared to the original simulated MRI with 0% noise and 0% RF using Root mean square error (RMSE) as a comparison metric. The RMSE was calculated for each MRI slice and the average RMSE for the entire datasets was used for comparison. Results