X Dai*, Y Lei, T Wang, J Zhou, J Roper, M McDonald, J Beitler, J Bradley, T Liu, X Yang, Emory Univ, Atlanta, GA

• 
  X Dai

• 
  

WE-D-TRACK 6-5 (Wednesday, 7/28/2021) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Purpose: Eliminating the labor-intensive, time-consuming, and observer-dependent manual delineation of organs-at-risk (OARs) is highly desired in radiotherapy treatment planning. This study aimed to tackle the challenges remaining in achieving reliable expert performance among current auto-delineation algorithms through developing a novel synthetic MRI-aided method.

Methods: MRI provides superior soft-tissue contrast compared to CT. By taking the complementary contrasts provided by CT and MRI into account, the accuracy of OAR delineation can be improved. In the study, synthetic MR images were firstly synthesized from CT images utilizing a cycle-consistent generative adversarial network. A novel model namely, mask scoring regional convolutional neural network, was implemented to obtaining delineation of organs through extracting and combining the features of CT and synthetic MRI. The performance of the proposed method was quantitatively assessed using both in-house and public datasets containing CT scans from head-and-neck (HN) cancer patients. Metrics including Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), mean surface distance (MSD), and residual mean square distance (RMS) were calculated to evaluate the outperformance of our proposed method against current state-of-the-art algorithms.

Results: A mean DSC of 0.77, mean HD95 of 2.90 mm, mean MSD of 0.89 mm, mean RMS of 1.44 mm were achieved by our proposed method across all of 18 OARs in our in-house dataset which contains images from 70 patients, which outperforms the current state-of-the-art algorithms by 6%, 16%, 25%, and 36%, respectively. A mean DSC of 0.86 was achieved for all 9 OARs on public dataset which has 48 patients, which is 6% better than the competing methods.

Conclusion: In this study, a synthetic MRI-aided deep learning framework for automated delineation of OARs in HN CT has been developed. The proposed method will have a high impact in rapid OAR contouring in routine HN cancer radiotherapy treatment planning.

Handouts

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email