Deep Learning Approaches for Deblurring Motion-Affected Brain MRI Scans
Abstract
During MRI acquisition, the blurred images caused by head motion present significant problems for neuroimage analysis. In this work, we create artificial blurred images using various kernels to simulate the real-world corrupted MRI scans and explore different machine learning methods to recover blurred images to the original sharp images. In particular, we train our models to take in the corrupted images as the input, and the corresponding original sharp images as the targets. We investigate and compare the performance of two deep learning-based approaches. The first method is a deep convolutional autoencoder (AE). The second method is the cycle structure convolutional autoencoder (cycle-AE). For the cycle-AE, we add another network after the AE architecture which takes the deblurred image as the input and learns to reproduce it back to the original corrupted image. This model is inspired by the Cycle GAN [] framework. Specifically, the first network strives to deblur the image, and the second tries to reblur the image back to the original. We conjecture that adding the second neural network will help to bring the deblurred image closer to the ground-truth. We first examine these two methods on 2D slices of the MRI scans and then apply similar techniques to 3D volumes. To solve the huge computational cost associated with working with 3D scans, we train our models using small cuboids of data randomly extracted from the whole scans. The performance evaluation is carried out based on the traditional image quality indexes, such as peak signal-to-noise ratio (PSNR), Root Mean Square Error (RMSE), Tissue Contrast Tscore (TCT) and Structural Similarity (SSIM) index. The results suggest that all our models can significantly deblur the images by visual observation. Furthermore, when images are partially damaged or the kernels which create the damaged image are more complicated, cycle-AE has a better performance than the baseline-AE.
Subject Area
Computer science|Medical imaging|Artificial intelligence
Recommended Citation
Wang, Tong, "Deep Learning Approaches for Deblurring Motion-Affected Brain MRI Scans" (2020). ETD Collection for Fordham University. AAI28086404.
https://research.library.fordham.edu/dissertations/AAI28086404
