Abstract Details

Title 3D MR image synthesis using unsupervised deep learning algorithm in MS patients

Topic Multiple Sclerosis

Presentation(s) P5 - Poster Session 5 (5:30 PM-6:30 PM)

Poster/Presentation
Number 15-025

Objective To generate synthetic 3D Fluid-Attenuated Inversion Recovery (FLAIR) images from acquired 3D T1 MR images of MS subjects.

Background Clinical MRI scans typically exhibit significant image heterogeneity between timepoints, representing a major barrier for quantitative post-processing analysis. Image synthesis methods can be used to generate standardized images from any clinical images without losing patient information. Generative adversarial networks (GANs) is a useful method for image-to-image translations in computer vision applications. We present an implementation of GANs for image-to-image translation of T1-to-FLAIR-weighted MR images and vice versa.

Design/Methods 400 MS T1/FLAIR pairs acquired at 3T were used to train our learning algorithm. Skulls were removed and each pair was co-registered for better pixel to pixel correspondence. We used Cycle-Consistent GAN, an unsupervised deep learning method, to generate 3D synthetic images. For validation purposes, three comparative methods were implemented in each slice; Mean Absolute Error (MAE), Mutual Information (MI) and Relative Error (RE). Comparisons were done between synthetic images as output and the acquired images as ground truth. To mitigate variations between input/output pairs, all slices were normalized by subtracting mean slice intensity divided by its standard deviation.

Results 63 additional MS cases with 3D T1 and 3D FLAIR images were used to validate the trained algorithm. MAE, MI, and RE were computed for each slice in each case and averaged altogether. MAE for all slices had a value of 0.8 as the average absolute difference between two images, average MI value was 1.28 which measures amount of mutual dependence between two images. RE map shows good match in normal appearing white/grey matter (NAWM/NAGM) while larger values are seen around lesion edges.

Conclusions

Synthetic images may represent a breakthrough for quantitative analysis of heterogeneous MS scans. Cycle-GAN shows excellent agreement in NAWM/NAGM against ground truth, while lesions may benefit from additional data incorporated into the model.

Authors/Disclosures
PRESENTER	No disclosure on file
Lokesh A. Rukmangadachar, MD	Dr. Rukmangadachar has nothing to disclose.
Daniel Pelletier, MD (Keck School of Medicine of USC)	Dr. Pelletier has received personal compensation in the range of $5,000-$9,999 for serving as a Consultant for Novartis. Dr. Pelletier has received personal compensation in the range of $10,000-$49,999 for serving as a Consultant for Sanofi Genzyme. Dr. Pelletier has received personal compensation in the range of $5,000-$9,999 for serving as a Consultant for Roche. Dr. Pelletier has received personal compensation in the range of $500-$4,999 for serving on a Scientific Advisory or Data Safety Monitoring board for Polpharma Biologics.

��ɫ��

��ɫ��