MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Hybrid Virtual MRI/CBCT Generation Using An Unsupervised Convolutional Neural Network (CNN) with Transfer Learning

Y Chen1*, F Yin2 , L Ren3 , (1) Duke University, Durham, NC, (2) Duke University Medical Center, Durham, NC, (3) Duke University Medical Center, Durham, NC

Presentations

(Sunday, 7/14/2019) 2:00 PM - 3:00 PM

Room: Stars at Night Ballroom 2-3

Purpose: On-board CBCT has very limited accuracy for localizing tumors in the soft tissue, such as liver tumor, due to its poor soft tissue contrast and various artifacts. This study aims to improve the localization accuracy of liver patients by generating hybrid virtual-MRI/CBCT.

Methods: The method involved two parts: (1). Liver contour generation in CBCT. Liver was contoured by physicians in the planning CT. An unsupervised CNN model was developed to register the contour deformation from CT to CBCT to generate liver contour in CBCT. (2). Hybrid virtual-MRI/CBCT generation. Liver was also contoured in the prior MRI. Surface-to-surface matching and the finite element method were used to deform liver volume from prior MRI to CBCT to generate virtual on-board MRI in the liver, which was then embedded into CBCT to generate hybrid images for both target localization and healthy tissue verification. In this study, we focus on demonstrating the feasibility of Part (1) of automatic liver contour generation in CBCT, since Part (2) is more straightforward and builds upon Part (1). For Part (1), a CNN deformable registration model was initially trained on brain MRI images and was then fine-tuned based on transfer learning using 14 liver patient CT and CBCT data. The trained model was used to register the deformation of a new patient with multiple day’s CBCT to generate liver contours in CBCT. Deformable registration accuracy using Velocity and our proposed CNN model were compared.

Results: CNN achieved equivalent or better deformation accuracy than Velocity. Mutual information, cross correlation and structural similarity index between deformed CT using CNN and CBCT are above 1.27, 0.97 and 0.91, respectively.

Conclusion: Preliminary results demonstrated the feasibility to use the unsupervised CNN to automatically generate liver contour in CBCT, which can then be used to generate hybrid virtual- MRI/CBCT for target localization.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH grant R01 CA-184173.

Keywords

Cone-beam CT, MRI

Taxonomy

IM/TH- RT X-ray Imaging: CBCT imaging/therapy implementation

Contact Email