Encrypted login | home

Program Information

Convolution Neural Network Based Deformable Image Registration

no image available
V Kearney

V Kearney*, S Haaf , A Sudhyadhom , T Solberg , UCSF Comprehensive Cancer Center, San Francisco, CA


MO-RAM-GePD-JT-5 (Monday, July 31, 2017) 9:30 AM - 10:00 AM Room: Joint Imaging-Therapy ePoster Theater

Purpose: Widespread skepticism in regards to the accuracy of deformable image registration (DIR), has limited the use of cone beam CT (CBCT) to CT DIR. Convolution neural networks (CNN) strategies have shown potential in the presence of shading, noise, and lighting variations, often present in CBCTs. Previously, deep learning DIR strategies have been used only in DIR objective functions. This approach relies on the whole network’s expression of a single differentiable score function, obfuscating much of the useful information in the lower convolutional layers and leading to inaccuracies. This study aims to demonstrate the improved accuracy and speed of using a novel graphics processor unit (GPU) based CNN-DIR algorithm to predict a set of points on a CBCT image given a set of points on a CT image.

Methods: 10 head and neck CBCT-CT image sets were used in this study. 100 random rigid shifts of 5 CBCT-CT image sets were used to construct the training set for the CNN. 10,000 random points were generated in each CT image set. CNN-DIR uses the CT image, CBCT image, and initial set of points to make a prediction about a deformed set of points. The deformed points are used to deformably map the CT to the CBCT, using sparse interpolation. The accuracy of the CNN DIR algorithm was evaluated using confined mutual information CMI, Canny-edge RMSE, and a synthetic ground truth.

Results: CNN-DIR showed improved accuracy compared to intensity-corrected Demons, and landmark-guided DIR, for all evaluation metrics and the synthetic ground truth. Using two Nvidia Geforce 1080 GPUs, the mean prediction time was 16 seconds.

Conclusion: Although CNN-DIR takes a fairly long time to train initially, each prediction event is computationally efficient enough for real-time clinical use. Additionally, CNN-DIR’s increased accuracy help unlock the true potential of on-board CBCTs

Contact Email: