J Harms*, Y Lei, S Tian, N Mccall, K Higgins, J Bradley, T Liu, X Yang, Winship Cancer Institute of Emory University, Atlanta, GA

• 
  J Harms

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

TU-E-TRACK 6-2 (Tuesday, 7/27/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: Radiation-induced heart disease is a post-treatment toxicity which directly impacts overall survival for many thoracic cancer patients. Dose to specific cardiac substructures can better predict post-treatment toxicity than whole heart dose. However, delineation of substructures during treatment planning is not feasible in most clinics. This study proposes a deep-learning-based framework for simultaneous, automatic generation of substructure contours which can be used to better understand the relationship between radiation dose and post-treatment toxicity.

Methods: The proposed mask-scoring regional convolutional neural network (MS-R-CNN) with attention gate consists of five subnetworks: backbone, regional proposal network (RPN), R-CNN, mask network with an attention gate, and mask-scoring network. Relying on CT alone, the backbone extracts multi-scale feature maps. The RPN utilizes these maps to detect the volume-of-interest (VOI) around all substructures, and the final three subnetworks extract structural information from these VOIs. An attention gate enforces a deep attention strategy which highlights informative features for accurate substructure segmentation. A retrospective study with 55 patients undergoing radiotherapy for lung cancer has been conducted to evaluate the proposed segmentation method. The whole heart and 15 cardiac substructures have quantitatively been accessed based on Dice score coefficients (DSC) and mean surface distances (MSD).

Results: The proposed method achieved average DSC for the whole heart, chambers, great vessels, coronary arteries, the valves of the heart of 0.96, 0.94, 0.93, 0.66, and 0.77 respectively. MSD between substructures segmented by the proposed method and the physicians’ manual contours were 1-5 mm for all substructures. After network training, all substructures and the whole heart can be segmented on new datasets in less than five seconds.

Conclusion: A deep-learning network was developed for automatic delineation of cardiac substructures based on CT alone. The tool has high potential for investigating cardiac substructure dose and treatment toxicities for improved cardiac sparing treatment planning.

Funding Support, Disclosures, and Conflict of Interest: This research is supported in part by the National Cancer Institute of the National Institutes of Health under Award Number R01-CA215718 and an Emory Winship Cancer Institute Pilot grant.

Handouts

Keywords

Not Applicable / None Entered.

Taxonomy

IM/TH- Image Segmentation Techniques: Machine Learning

Contact Email