MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Automatic CBCT-Based Multi-Organ Segmentation Using Deep-Learning for Pancreatic Adaptive Radiotherapy

Y Liu1*, Y Lei2, J Janopaul-naylor3, J Harms4, T Wang5, W Curran6, P Patel7, T Liu8, X Yang9, (1) Emory University, Atlanta, GA, (2) Emory University, Atlanta, GA, (3) ,,,(4) Emory University, Atlanta, GA, (5) Emory University, Atlanta, GA, (6) Winship Cancer Institute, Atlanta, GA, (7) Emory University, Atlanta, GA, (8) Emory Univ, Atlanta, GA, (9) Emory University, Atlanta, GA

Presentations

(Sunday, 7/12/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: In addition to the current use for target localization and patient setup, CBCT has the potential to be further employed for adaptive radiotherapy (ART) by evaluating the dose to be delivered to organs-at-risk (OARs). One of the major challenges for CBCT-guided ART is rapid and accurate multi-organ segmentation, which is required for real-time OAR DVH evaluation. This study aims to develop a deep-learning-based method for CBCT-based automatic OAR segmentation aided by synthetic CT (sCT) for pancreatic ART.


Methods: Forty patients who received RT for pancreatic carcinoma had their OARs manually contoured on CBCT images. A cycle-consistent generative adversarial network was used to generate high-quality CBCT-based sCTs. Subsequently, a fully convolutional network used these sCTs and manual contours for training, and output a multi-channel binary mask for all OARs. A deep attention strategy was used to highlight the salient features that accurately represent the different organs. Five-fold cross-validation was used to evaluate the method. Dice similarity coefficient (DSC), mean surface distance (MSD) and percentage volume difference (PVD) were used to quantify the differences between the manual and automated contours.


Results: The mean DSC, MSD and PVD are 0.80±0.20, 1.95±2.37mm and 10.81±9.28% for duodenum, 0.87±0.05, 1.33±0.58mm and 4.06±4.43% for small bowel, 0.88±0.07, 2.10±2.21mm and 4.68±6.16% for large bowel, 0.91±0.08, 1.59±1.92 mm and 6.09±7.37% for stomach, 0.93±0.05, 2.90±4.77 mm and 4.31±5.68% for liver, 0.92±0.05, 2.05±4.71 mm and 5.20±5.35% for left kidney, 0.93±0.04, 2.09±3.40mm and 2.58±5.06% for right kidney and 0.88±0.07, 0.75±0.50 mm and 7.20±9.58% for spinal cord. All OAR segmentation is achieved in a few seconds.


Conclusion: This study demonstrates that a novel sCT-aided deep-learning-based method is capable of rapidly and accurately segmenting OARs for pancreatic ART. This segmentation tool warrants further development of a CBCT-guided ART workflow for pancreatic cancer.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email