Click here to

Session: AI Applications in Image Guided Adaptive Radiation Therapy [Return to Session]

CBCT-Based Prostate and Organs-At-Risk Segmentation Using Deep Attention Network

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


TH-E-TRACK 4-6 (Thursday, 7/29/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: To propose a deep attention network for cone beam CT (CBCT)-based multi-organ segmentation for facilitating CBCT-guided adaptive radiotherapy.

Methods: Both the CBCT and synthetic MRI (sMRI) were used for the segmentation of soft tissue including the prostate, bladder, rectum, and the bones including left femoral head (LFH) and right femoral head (RFH). The synthetic MRI was generated using trained CycleGAN networks. The CBCT and sMRI were separately processed using two U-Nets for feature extraction. The CBCT has better ability in bony structure segmentation while the sMRI has better ability for soft tissue discrimination. It is important to effectively combine the feature maps learnt from both the CBCT and sMRI. To achieve this, a novel deep attention strategy was proposed to integrate the CBCT feature maps with the sMRI feature maps prior to final segmentation by a convolutional neural network (CNN). The deep attention strategy enables the network to focus on salient features that highlight the target organs. The network was trained and tested using 50 patients’ datasets using five-fold cross validation. For evaluation, the dice similarity coefficient (DSC) and mean surface distances (MSD) were calculated for the target organs.

Results: Our results show that the proposed method was able to segment all five organs simultaneously with a high accuracy. The DSCs and MSDs are 0.90±0.06 and 0.94±0.91 mm, 0.94±0.04 and 0.87±0.76 mm, 0.90±0.03 and 0.88±0.73 mm, 0.96±0.04 and 1.16±1.34 mm, as well as 0.95±0.03 and 1.28±1.43 mm for prostate, bladder, rectum, LFH and RFH, respectively. This proposed sMRI-aided CBCT segmentation outperforms CBCT-only based segmentations with better agreement with physicians’ manual contours.

Conclusion: We have developed a CNN network with deep attention strategy to segment prostate and multiple OARs on CBCT images. The accurate sMRI-aided organ delineation warrants further development of a CBCT-guided adaptive radiotherapy workflow for prostate cancer.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email