Click here to


Are you sure ?

Yes, do it No, cancel

Cardiac Substructure Segmentation with Deep Learning for Improved Cardiac Sparing

E Morris1,2*, A Ghanem1,3, M Dong4, H Emami4 , M Pantelic5, E Walker1, C Glide-Hurst1,2, (1) Department of Radiation Oncology, Henry Ford Cancer Institute, Detroit, Michigan, (2) Department of Radiation Oncology, Wayne State University School of Medicine, Detroit, Michigan, (3) Department of Clinical Oncology, Alexandria University, Alexandria, Egypt, (4) Department of Computer Science, Wayne State University School of Medicine, Detroit, Michigan, (5) Departments of Radiology, Henry Ford Cancer Institute, Detroit, Michigan


(Monday, 7/15/2019) 1:45 PM - 3:45 PM

Room: Stars at Night Ballroom 2-3

Purpose: Radiation dose to cardiac substructures is related to radiation-induced heart disease. However, substructures are not considered in radiation therapy planning (RTP) due to poor visualization on CT. To address this, we developed a novel deep learning pipeline leveraging the soft tissue contrast of MRI coupled with CT for state-of-the-art substructure segmentation requiring only non-contrast CT inputs.

Methods: Thirty-two left-sided whole-breast cancer patients underwent cardiac T2 MRI and CT-simulation. A rigid cardiac-confined MR/CT registration enabled ground-truth delineations of 12 substructures (chambers, great vessels, coronary arteries, etc.). Paired MRI/CT data for 25 patients were placed into separate image channels to train a 3-dimensional Neural Network (3D U-Net) using the entire 3D image and all substructures simultaneously. A Dice-weighted multiclass loss function was utilized along with deep supervision. Segmentation results were assessed pre/post augmentation (scaling, rotation, flipping, and translation). Results for 11 test patient CTs (7 unique patients) were compared to ground-truth and our previously developed multi-atlas method. Prediction versus ground-truth evaluation was performed via Dice similarity coefficient (DSC) and mean distance to agreement (MDA).

Results: The model stabilized in ~17 hours after 200 epochs (training error<0.001). Augmentation increased DSC ~5% across substructures. Deep learning provided accurate segmentations for chambers (DSC: 0.87±0.03), great vessels (DSC: 0.85±0.03), and pulmonary veins (DSC: 0.71±0.03). Combined DSC for coronary arteries was 0.46±0.09. MDA across 12 substructures was <2.0mm (great vessel MDA: 1.11±0.2mm). Compared to our previous multi-atlas method, MDA improved ~1.4mm while DSC increased 3-7% (chambers) and 23-35% (coronary arteries). For four test CTs, 3D U-Net yielded left main coronary artery contours, whereas the atlas-based segmentation failed. Contour generation takes ~5 seconds (<1% of multi-atlas time).

Conclusion: These promising results suggest deep learning offers major efficiency and accuracy gains for cardiac substructure segmentation offering high potential for rapid implementation into RTP for improved cardiac sparing.

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 National Cancer Institute of the National Institutes of Health under Award Number R01CA204189. Additionally, this work was partially supported by US National Science Foundation (NSF) under grant CNS-1637312.


Heart, Radiation Therapy, Segmentation


IM/TH- image segmentation: Multi-modality segmentation

Contact Email