Click here to


Are you sure ?

Yes, do it No, cancel

BEST IN PHYSICS (MULTI-DISCIPLINARY): A Deep Learning Cardiac Substructure Pipeline for MR-Guided Cardiac Applications

E Morris1,2*, A Ghanem3, M Dong4, S Zhu1, M Pantelic5, C Glide-Hurst1,2, (1) Department of Radiation Oncology, Henry Ford Cancer Institute, Detroit, MI (2) Department of Radiation Oncology, Wayne State University School of Medicine, Detroit, MI (3) Department of Clinical Oncology, Alexandria University, Alexandria, Egypt (4) Department of Computer Science, Wayne State University, Detroit, MI (5) Department of Radiology, Henry Ford Cancer Institute, Detroit, MI


(Monday, 7/13/2020) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: Recent evidence suggests that radiation doses to cardiac substructures are more strongly linked to late cardiac morbidities than commonly used whole-heart dose metrics. MR-guided radiation therapy presents a unique opportunity to visualize sensitive cardiac substructure location during tumor localization and for daily treatment, which allows the potential for enhanced cardiac sparing. We have established a deep learning (DL) model for rapid cardiac substructure segmentation based on low field MRI.

Methods: Twenty-three patients were retrospectively evaluated who underwent thoracic radiation therapy with a 0.35T MR-guided linac. Twenty-one patients were treated under breath-hold conditions (17-25s, 14 end-inhalation and 7 end-exhalation, 1.5x1.5x3 mm³) and two under free breathing conditions (3-minutes, 1.5 mm³ isotropic resolution). Two radiation oncologists generated ground-truth segmentations of 12 cardiac substructures (chambers, great vessels, coronary arteries, etc.) using 0.35T MRI for 114 MRIs (4-5/patient). Eighteen patients (n=90) were used to train a three-dimensional U-Net with a Dice-weighted focal loss function to manage differences in substructure size and complexity. The remaining five patients were a holdout test set to assess mean distance to agreement (MDA) and Dice similarity coefficient (DSC) to ground-truth.

Results: The model stabilized after training for 340 epochs (training error <0.001) which took ~99 hours to complete. Segmentation results varied slightly depending on scan duration (< ±0.03 change in DSC). DL provided accurate segmentations for the chambers (DSC=0.85±0.03), great vessels (DSC=0.81±0.04), and pulmonary veins (DSC=0.71±0.03). DSC for the coronary arteries was 0.42±0.07. MDA across all 12 substructures was <3.0 mm. Wilcoxon signed ranks test revealed no cardiac substructures had significant differences in volume between DL segmentations and ground-truth (p>0.05). Substructure contour generation for a new patient input took ~15 seconds.

Conclusion: This work is a critical step in providing rapid and reliable cardiac substructure segmentations that may be used to improve cardiac sparing in MR-guided radiation therapy.

Funding Support, Disclosures, and Conflict of Interest: The submitting institution holds research agreements with Philips Healthcare, ViewRay, Inc., and Modus Medical. Research partially supported by the NCI/NIH, Award R01CA204189. The PI is on the Philips Healthcare Advisory Board. This work was also partially supported by US National Science Foundation (NSF) under grant CNS-1637312.


Heart, Image-guided Therapy, MRI


IM/TH- image Segmentation: MRI

Contact Email