Y Liu*, Y Lei, T Wang, Y Fu, P Patel, A Jani, J Bradley, T Liu, X Yang, Emory Univ, Atlanta, GA

• 
  Y Liu

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

TH-E-TRACK 4-7 (Thursday, 7/29/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: Fast and accurate segmentation of target and organs-at-risk (OARs) is a key requirement in adaptive radiotherapy (ART). In the scenario of CT-on-rails-based ART, the absence of MRI may lead to inaccurate prostate delineation. We propose to develop a CT-only, synthetic MRI (sMRI)-aided dual networks for rapid and accurate multi-organ segmentation for prostate ART.

Methods: We propose a dual network method that combines the bony structure information provided by CT and superior soft-tissue information provided by sMRI for effective pelvic multi-organ segmentation. Cycle-consistent adversarial networks (CycleGAN) were used to generate CT-based sMRI. Independent features are extracted from CT and sMRI via the dual networks separately: the first network is used to extract features from CT that represent bony structures, and the second network is used to explore features from sMRI that represent the soft tissues. Deep supervision is used to force the extracted CT and sMRI feature maps from hidden layers of each pyramid level to be semantically discriminative. Then, the comprehensive feature maps from corresponding pyramid levels of the first and second pyramid networks are combined and processed via several deconvolution layers and attention gates. The attention gates further highlighted the informative elements of combined feature maps to well differentiate different organ tissues. The proposed method was trained and evaluated using a cohort of 140 prostate patients.

Results: The Dice similarity coefficients and mean surface distances between our results and ground truth were 0.95±0.05, 1.16±0.70mm; 0.88±0.08, 1.64±1.26mm; 0.90±0.04, 1.27±0.48mm; 0.95±0.04, 1.08±1.29mm; and 0.95±0.04, 1.11±1.49mm for bladder, prostate, rectum, left and right femoral heads, respectively. The mean center of mass distances was within 3 mm for all organs.

Conclusion: We demonstrated the feasibility of sMRI-aided dual networks for multi-organ segmentation on pelvic CT images. This method could reduce clinical workload and facilitate the development of CT-on-rail-based adaptive radiotherapy.

Handouts

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email