S2101
Physics - Image acquisition and processing
ESTRO 2026
Digital Poster 5039 Improved Abdominal Gas Cavity Definition in Synthetic CT for MRI-Only Simulation Imaging with MRI-Guided Gastrointestinal Radiotherapy Braian Adair Maldonado Luna 1 , Kamal Singhrao 2 , Benito de Celis 1 , Gerardo Uriel Pérez Rojas 1 , René Eduardo Rodríguez Pérez 1 1 Faculty of Mathematical Physical Sciences, Benemérita Universidad Autónoma de Puebla, Puebla, Mexico. 2 Department of Radiation Oncology, Brigham and Women's Hospital, Dana-Farber Cancer Institute, Harvard Medical School, Boston, USA Purpose/Objective: Integration of MRI-only simulation with MRI-guided radiotherapy (MRgRT) can potentially improve online adaptive planning efficiency. However, inaccurate abdominal gas cavity definition in deep-learning- (DL)- derived synthetic CT images, which are used for electron density calculations in MRI-only simulation workflows, hampers the use of sCT for high- throughput MRgRT treatment sites such as the abdomen [1]. A cause of inaccurate abdominal gas cavity definition is due to poor deformable image registration (DIR) of gas cavities during sCT model training [2]. Here, we validate a novel sCT generation method, which eliminates the need for DIR mediation during model training and evaluate the gas cavity definition accuracy. We also quantify the differences in gas cavity definition using our sCT method to the standard single stage MRI-CT DL modeling training method. Material/Methods: SCT images were generated using a two-stage generative-DL approach: where the first stage involves a generative DL model to auto-segment tissues in MRI and the second generative DL model converts an auto- segmentation volume to a synthetic CT. Preprocessing included segmenting tissue-specific labels—such as gastrointestinal (GI) contours, gas cavities, soft tissues, and bone—through manual contouring and AI-based autocontouring, defined independently in MRI and CT. Both stages were trained using paired images from 12 abdominal cancer patients [3], 9 for training and 3 for testing. The tissue-specific segmentation accuracy for the single-stage and two-stage methods was evaluated using the Dice-Sørensen coefficient in both the sCT images and the MRI-derived segmented maps, to assess anatomical accuracy. Additionally, a tissue- specific Dice-Sørensen coefficient was calculated for gas cavities, identified as the primary area of interest. Results: The global sCT Dice-Sørensen coefficient for the two- stage method and the standard single-stage method were 0.99±0 and 0.91±0.01 respectively. The gas cavity Dice-Sørensen coefficient for the two-stage method
accuracy. We conducted a retrospective review (IRB- 19-2033) of 238 patients that received 5DCT, 107 of which had tumors contoured in the lungs. Breathing statistics, including session breathing amplitude, period and regularity, the accuracy of the surrogate, and the motion model residuals were evaluated. For patients with tumors, we focused an additional evaluation on the tumor locations, using the motion model to alter each of the 25 scans to the reference scan geometry and comparing the reference scan to the reconstructed scans. Results: The surrogate accuracy showed that 94% of the patients had a surrogate accuracy of better than 8% of the session breathing amplitude. Session breathing amplitude was a mean of 19.9±7.7mm (range 6.2mm- 45.9mm). The 90th percentile model error (within the lungs) ranged from 0.8mm-6.0mm with 90% of patients having an error less than 2.9 mm. The mean tumor motion model RMS residual was 1.1mm (range 0.3 mm-2.8 mm). Tumors that moved >1 cm and >2 cm had less than 20% and 10% errors, respectively, compared against the motion magnitude. The mean tumor position accuracy of the model was 1.1±0.7 mm (range 0.2 mm-3.3 mm, See Figure). The maximum single scan error ranged from 0.5 mm-6.7 mm with the outlier at 12.6 mm proving to be imaged during a clearing breath. 83% of the imaged tumors had their worst scan error less than 3 mm.
Conclusion: 5DCT was shown to be both precise and accurate with respect to modeling human breathing motion. Work continues to address highly irregular breathing patients and improving the clinical workflow. References: 1. Low DA, Parikh PJ, Lu W, et al. Novel breathing motion model for radiotherapy. Int J Radiat Oncol Biol Phys. 2005;63(3):921-929.2. Mattias P. Heinrich, M. Jenkinson, M. Brady and J.A. Schnabel, MRF-Based Deformable Registration and Ventilation Estimation of Lung CT. IEEE Transactions on Medical Imaging 2013, 32(7): 1239-1248 Keywords: 5DCT, 4DCT, Breathing Motion Imaging
Made with FlippingBook - Share PDF online