Deep Learning-Based Cardiac Chamber Segmentation in Magnetic Resonance-Guided Adaptive Radiation Therapy

Authors

Xinru Chen
Yao Ding
Julius Weng
Carol C Wu
Yao Zhao
Angela Sobremonte
Mustefa Mohammedsaid
Zhan Xu
Xiaodong Zhang
Joshua S Niedzielski
Sanjay S Shete
Laurence E Court
Zhongxing Liao
Jihong Wang
Ergys Subashi
Percy P Lee
Jinzhong Yang

Publication Date

9-1-2025

Journal

Advances in Radiation Oncology

DOI

10.1016/j.adro.2025.101845

PMID

40808698

PMCID

PMC12344783

PubMedCentral® Posted Date

7-4-2025

PubMedCentral® Full Text Version

Post-print

Abstract

Purpose: Accurate cardiac chamber segmentation is crucial for improving cardiac sparing in magnetic resonance (MR)-guided adaptive radiation therapy, especially in patients at risk for radiation-induced cardiotoxicity. Here, we developed and evaluated automatic segmentation models for cardiac chambers that use daily MR images acquired on a 1.5-T MR-Linac system.

Methods and materials: Twenty healthy volunteers underwent daily MR scanning on a 1.5-T MR-Linac, with 2 radial sequences: T2/T1 3DVaneXD balanced fast field echo with spectral attenuated inversion recovery (bFFE-SPAIR) and T1 3DVaneXD mDixon. Three flip angles were tested for each sequence to determine optimal image quality for chamber segmentation. Full-resolution 3D nnU-Net models were trained for the following: (1) bFFE-SPAIR (bFFE model); (2) T1 mDixon (mDixon model); and (3) both sequences (hybrid model). Models were evaluated based on Dice similarity coefficient (DSC) and mean surface distance against manual contours. Clinical acceptance of the automatic segmentation was assessed with a 5-point Likert scale. An in-silico planning study was performed to assess cardiac chamber sparing during plan adaptation.

Results: The average contrast-to-noise ratios in bFFE-SPAIR were 8.7 (20°), 34.2 (50°), and 37.3 (80°); for T1 mDixon, these values were 3.6 (5°), 5.9 (10°), and 4.9 (20°). The bFFE model achieved the highest segmentation performance (average DSC 0.85 ± 0.05 and mean surface distance 2.2 ± 0.6 mm). The T1 mDixon sequence, despite lower contrast-to-noise ratios, provided similar segmentation accuracy (DSC 0.83 ± 0.06). A hybrid model combining both sequences showed no significant improvement over the bFFE model. Clinical evaluation indicated that 95% of the autosegmented contours from the bFFE model were acceptable for clinical use (score ≥4). Adaptive plan greatly reduced individual cardiac chamber dose while maintaining similar target coverage.

Conclusions: This study demonstrated the feasibility of using bFFE-SPAIR and T1 mDixon sequences to accurately segment cardiac chambers on a 1.5-T MR-Linac. These models offer potential for improved cardiac sparing in MR-guided adaptive radiation therapy.

Published Open-Access

yes

Recommended Citation

Chen, Xinru; Ding, Yao; Weng, Julius; et al., "Deep Learning-Based Cardiac Chamber Segmentation in Magnetic Resonance-Guided Adaptive Radiation Therapy" (2025). Faculty, Staff and Student Publications. 4905.
https://digitalcommons.library.tmc.edu/uthgsbs_docs/4905