MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Multi-Organ Segmentation in Head-And-Neck CT Using Mask Scoring Regional Convolutional Neural Network (MS-RCNN)

Y Lei, X Dai*, Y Liu, T Wang, X Jiang, J Beitler, W Curran, T Liu, X Yang, Department of Radiation Oncology and Winship Cancer Institute, Emory University, Atlanta, GA

Presentations

(Monday, 7/13/2020) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: Organs-at-risk (OARs) delineation is an important, but labor intensive and time-consuming procedure during the treatment planning process for head-and-neck (HN) cancer radiotherapy. In this study, a fully automatic multi-organ segmentation method based on advanced deep-learning algorithm is implemented and evaluated to facilitate radiation therapy for HN cancers.

Methods: A mask scoring regional convolutional neural network (MS-RCNN) aided by synthetic MRI (sMRI) has been developed for automatic contouring OARs in HN CT images. Prior to training, sMRI were firstly generated from CT images using a pre-trained cycle-consistent generative adversarial network (CylceGAN) based model. The sMRI is then concatenated with CT as a two-channel input to a mask scoring R-CNN model, in which residual learning architecture ResNet101 is used as backbone while pyramid feature collection and regional proposal network being adopted to predict the location and size of each volume-of-interest. Binary masks obtained by a fully convolution network using pyramid features and predicted ROIs, are then used to supervise the segmentation of OARs. CT images of 40 HN cancer patients were manually contoured by experienced physicians and then were then used as annotated data sets to evaluate the proposed method.

Results: The mean Dice similarity coefficients (DSC) of spinal cord, oral cavity, mandible, larynx, pharynx, esophagus, right parotid, and left parotid are 0.85±0.02, 0.89±0.02, 0.88±0.03, 0.83±0.08, 0.81±0.04, 0.82±0.06, 0.79±0.05, and 0.82±0.05, respectively. All OAR segmentation can be achieved in a few seconds.

Conclusion: In this study, an advanced deep learning-based fully automatic multi-organ segmentation algorithm has been investigated for OARs delineation in CT images of HN patients. The results show the superb performance in term of accuracy in multi-organ contouring, and further demonstrate the potential of adopting our proposed approach in current clinical workflow for HN cancers radiotherapy.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email