Click here to

Session: Multi-Disciplinary: Image Guidance: Cone-Beam CT [Return to Session]

Artificial Intelligence (AI) Based Missing Tissue Generated Synthetic CT Preparation for Planning in Limited Field of View (LFOV) MR-Only Simulation

S Kim1*, S Kim2, (1) Catholic University of Korea, Seoul, ,KR, (2) Virginia Commonwealth University, Glen Allen, VA


SU-IePD-TRACK 3-7 (Sunday, 7/25/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: MR-only simulations often suffer from limited field of view (LFOV), resulting in incomplete body outlines. A novel approach of generating such missing tissue using machine learning and optical imaging has been conceived.

Methods: The method consists of 3 steps. 1) Patient body outline in interest is obtained using an optical imaging method. 2) An MR image set in limited FOV (MR-in-LFOV) is converted to a synthetic CT image set in limited FOV (syn-CT-in-LFOV). This task can be accomplished using any one or combination of existing methods such as an intensity based approach, an atlas based approach and a machine learning based approach. 3) The syn-CT-in-LFOV is expanded to a synthetic CT image set in full FOV (syn-CT-in-FFOV) using a machine learning based missing body generation algorithm. The main goal of step 3 (i.e., generation of missing part of image) has not been proven and we focused on its realization using regular CT images in this study. For proof of the concept, we built a model using 10,005 training images taken from the cancer imaging archive (TCIA) and performed tests with other data sets not used in training. All training and evaluation were performed on a 64‐bit Windows 10 Enterprise workstation. Computing hardware included an Intel Xeon W3520 quad‐core CPU, 32 GB RAM, and two Nvidia GeForce GTX 1080 Ti graphic cards with a total 7168 cores and 22GB GPU RAM.

Results: The completion network was trained for 250 iterations and the entire training procedure took roughly 5 days. The model provided missing tissues in tested cases.

Conclusion: It is demonstrated that the approach introduced can realistically complete CT images despite large occluded areas. This approach is also expected to reduce geometrical distortion that is dominant at periphery in typical MR images



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email