Click here to


Are you sure ?

Yes, do it No, cancel

Deep Learning for Rapid Deformable Image Registration of Liver CT Scans

B Anderson*, G Cazoulat , E Lin , B Odisio , K Brock , UT MD Anderson Cancer Center, Houston, TX


(Tuesday, 7/16/2019) 1:45 PM - 3:45 PM

Room: Stars at Night Ballroom 2-3

Purpose: To develop a deformable registration technique that has the accuracy of a biomechanical model-based algorithm with the speed required for on-line image guidance.

Methods: Eighty-three patients treated for colorectal liver metastasis with image guided microwave ablation were retrospectively evaluated under an IRB protocol, with forty-nine having been analyzed to date. For each case, a contrast enhanced CT scan (CECT) was obtained prior to, and at the completion of, the microwave ablation treatment. A biomechanical model-based deformable image registration method (DIR) algorithm was used to deform the pre-treatment CECT onto the post-treatment CECT. The resulting deformation vector field (DVF) was represented as a 3 channel array: x(left-right), y(anterior-posterior), and z(superior-inferior) directions. Liver focus regions were identified automatically on the CECTs and used as model inputs. 3D-UNets were developed and evaluated with variations in residual connections, depth, filter sizes, and filter number. The network was driven by minimizing the mean square distance (MSD) error. Leave-nine-out cross validation was used to create 2 training/testing sets of 40/9 image sets, respectively. The optimal model was selected based on a smallest MSD error.

Results: The optimal algorithm had a median difference in the x, y, and z directions of 1.4 mm, 1.7 mm, and 2.0 mm. The difference in Dice scores between the prediction and ground-truth DVF was µ:0.06, σ:0.04. The 3D-UNet created the DVF in 10 seconds on an 8GB GPU, compared to several minutes from the biomechanical DIR.

Conclusion: An AI-based algorithm was developed to mimic the results of a biomechanical model-based DIR algorithm. Using a small initial training set, the 3D-UNet achieved a median accuracy of less than 2.0 mm in each direction, compared to the full biomechanical model, in a fraction of the time, enabling it’s use in on-line image guidance. The additional training sets will likely improve the performance.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email