X Yang¹*, J Huang² , Y Lei³ , S Tian⁴ , K Higgins⁵ , J Beitler⁶ , D Yu⁷ , X Jiang⁸ , T Liu⁹ , W Curran¹⁰ , Y Fang¹¹ , (1) Emory University, Atlanta, GA, (2) New York University, New York, NY, (3) Emory University, Atlanta, GA, (4) Emory University, Atlanta, GA, (5) Emory University, Atlanta, GA, (6) Emory University, Atlanta, GA, (7) Emory University, Atlanta, GA, (8) Emory University, Atlanta, GA, (9) Emory Univ, Atlanta, GA, (10) Winship Cancer Institute, Atlanta, GA, (11) New York University, New York, NY

• 
  X Yang

(Sunday, 7/29/2018) 2:05 PM - 3:00 PM

Room: Room 205

Purpose: Accurate and fast segmentation of organs-at-risks (OARs) is the key step for efficient planning of radiation therapy for cancer treatment. The purpose of this work is to develop a deep learning-based method to automatically segment multiple OARs in chest CT image for radiotherapy treatment planning.

Methods: We propose an adversarial training strategy to train deep neural networks for the segmentation of multiple organs in chest CT image simultaneously. The proposed design of adversarial networks, called Med-A-Nets, jointly trains a set of convolution neural network (CNN), recurrent neural network (RNN) and an adversarial discriminator for the robust segmentation of multiple organs in CT image. More specifically, the generator, composed of CNN and RNN, produces image segmentation map of multiple organs in CT images by finding the optimal alignment with the ground truth segmentation map, whereas the discriminator, learned from ground truth segmentation map, tends to penalize the segmentation map produced by the generator. The generator and discriminator compete to each other in an adversarial learning process until the equivalence point is reached to produce the optimal segmentation map of multiple organs. Our segmentation results were compared with manual segmentation (ground truth) to quantitatively evaluate segmentation accuracy using Dice similarity coefficient (DSC) and intersection-over-union (IOU) metric

Results: This segmentation technique was validated on segmentation of left and right lungs, spinal cord, esophagus, and heart using 15 chest CT images. The mean DSC of all organs was 0.912 and IOU was 0.835, which indicates that the automatic segmentation method works well and could be used for radiotherapy treatment planning.

Conclusion: We have investigated a novel deep-learning-based approach with generative adversarial network strategy to segment multiple OARs in chest CT images and demonstrated its feasibility and reliability. This automated segmentation tool has a potential to improve efficiency of the current radiotherapy planning.

Keywords

Not Applicable / None Entered.

Taxonomy

IM- CT: Segmentation

Contact Email