(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: To develop and evaluate deep learning (DL) based auto-segmentation of cardiac substructures from non-contrast planning CT images in patients undergoing breast cancer radiotherapy.
Methods: Nine substructures including Aortic Valve(AV), Left Anterior Descending(LAD), Tricuspid Valve(TV), Mitral Valve(MV), Pulmonic Valve(PV), Right Atrium(RA), Right Ventricle(RV), Left Atrium(LA) and Left Ventricle(LV) were manually delineated by physician on non-contrast CT images of 50 patients with breast cancer. The image/label pairs of 30 subjects were used to train a 3D deep neural network while the remaining 20 were used for validation and testing. Our custom-designed network was based on a ResNet design. Random non-rigid and rigid transformations were used to augment the dataset. We investigated multiple loss functions including Dice Similarity Coefficient (DCS), Cross-entropy (CE), Euclidean loss as well as combinations of them. Without post-processing, the predicted label maps were compared to ground truth label via DSC and mean and 90th percentile symmetric surface distance (90th-SSD).
Results: When using combination of CE and DSC (CE-D) as loss function, DL achieved a mean DSC=0.80±0.03 for chambers and a mean DSC=0.53±0.06 for smaller structures (AV, LAD, PV, MV, TV). The mean and 90th-SSD for chambers is 1.91±0.26 mm and 4.60±0.69 mm and is 2.15±0.75 mm and 5.10±1.44 mm for smaller structures. Modified Dice (MD) loss function performs best in segmenting chambers when evaluating with 90th-SSD (4.45±80 mm), but worse in segmenting LAD (90th-SSD=8.66±2.88 mm), while CE-D as loss function provides best segmentation on smaller structures. The execution time for segmenting each patient is around 2.1s.
Conclusion: DL provides a fast and accurate segmentation of large cardiac substructures in non-contrast CT images, although performance for smaller structures was dependent on choice of loss function. Evaluation of clinical acceptability and integration into clinical workflow are pending.
Funding Support, Disclosures, and Conflict of Interest: GH reports honoraria / travel costs, Varian; research grants: Varian Medical Systems, Viewray, Siemens. MAT reports honoraria, ViewRay. CGR reports stock / leadership, Radialogica; consulting: Varian, AstraZeneca, EMD Serono; research grants: Varian, Elekta, Merck; travel costs: Siemens. JK reports: research grants: Varian; travel costs and honoraria: Varian
Keywords
Segmentation, Image Processing, Heart
Taxonomy
IM/TH- Image Segmentation Techniques: Machine Learning
Contact Email
