2451 - Feasibility of Cardiac Chamber Auto-Segmentation and Dose Optimization for MR-Guided Online Adaptive Radiotherapy of Central Lung Cancer

X. Chen¹, J. Weng², Y. Ding¹, A. Sobremonte², M. Mohammedsaid², Y. Zhao¹, X. Zhang¹, J. Niedzielski¹, S. S. Shete³, L. E. Court¹, Z. Liao², J. Wang¹, P. Lee⁴, and J. Yang¹; ¹Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, ²Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, ³The University of Texas MD Anderson Cancer Center UTHealth Houston Graduate School of Biomedical Sciences, Houston, TX, ⁴Department of Radiation Oncology, Lennar Foundation Comprehensive Cancer Center, City of Hope National Medical Center, Irvine, CA

Purpose/Objective(s): To investigate the feasibility of accurate auto-segmentation of cardiac chambers from MR images for cardiac dose sparing in MR-based online adaptive planning for central lung cancer stereotactic body radiotherapy (SBRT). Materials/

Methods: Images of 20 volunteers were acquired on a 1.5T MR-Linac system with 2 MR sequences for plan adaptation covering the entire thorax: T1 3DVaneXD mDixon (T1-mDixon) and balanced FFE 3DVaneXD SPAIR (bFFE-SPAIR). In both sequences, flip angles (FA) were optimized to achieve the best signal-to-noise ratio (SNR) and contrast-to-noise-ratio (CNR) of the cardiac chambers with the consideration of specific absorption rate (SAR) and acquisition time. MR images with optimal parameters were used by a radiation oncologist to delineate the whole heart and four chambers. Three deep-learning segmentation models based on nnU-Net were trained using: (1) the T1-mDixon sequence; (2) the bFFE-SPAIR sequence, and (3) both sequences Segmentation results were assessed using the Dice Similarity Coefficient (DSC). An in-silico adaptive planning was performed using a real patient plan with a treatment target in left lung near left atrium and a prescription of 70 Gy in 10 fractions. Adaptive plan was assessed by considering optimizing dose to heart chambers.

Results: Optimal FAs were selected as 10° for T1-mDixon and 50° for bFFE-SPAIR to improve chamber visualization while ensuring patient comfort. The acquisition time of T1-mDixon and bFFE-SPAIR was 6 minutes and 5.4 minutes respectively. The bFFE-SPAIR images showed better contrast than T1-mDixon in differentiating heart chambers from surrounding tissues during manual delineation. The average DSC values of the three deep learning models were 0.85±0.07, 0.84±0.07, and 0.85±0.08, demonstrating accurate segmentation results. In in-silico planning, by optimizing dose to left atrium and right ventricle, maximum dose to left atrium was reduced from 53.0Gy to 34.3Gy; and V10Gy for right ventricle was reduced from 5% to 2.6% to achieve V10Gy<4% goal. Other dosimetric parameters are similar before and after chamber dose optimization.

Conclusion: We demonstrated the feasibility of accurate segmentation of heart chambers from MR sequence for online adaptive planning in a 1.5 Tesla MR-Linac system to achieve improved heart chamber dose sparing for central lung cancer SBRT. Table 1. Dice Similarity Coefficient of three segmentation models

	Whole Heart	Left Atrium	Right Atrium	Left Ventricle	Right Ventricle
T1-mDixon	0.93±0.03	0.76±0.04	0.82±0.03	0.89±0.03	0.83±0.06
bFFE-SPAIR	0.93±0.02	0.78±0.03	0.79±0.03	0.89±0.03	0.83±0.03
Both	0.94±0.02	0.74±0.06	0.82±0.03	0.89±0.03	0.84±0.05