Optical Flow-Based Image Registration In Flair MRI

Medical imaging is one of the most common areas of computer vision research and algorithm development. FLAIR-MRI is particularly useful in highlighting damaged and necrotic tissue in brain images due to high contrast and resolution. Image registration is a method of warping images to the same geometric space to quantify tissue changes with accuracy. With advances in deep-learning via convolutional neural networks, complex problems can now move closer to some semblance of a solution with purpose-built and domain specific models. To overcome the non-learnable nature of current registration algorithms, ideas are adapted from video processing solutions of calculating optical flow between temporally spaced frames using unsupervised CNN-based methods to warp moving medical images to a fixed image space. The proposed total network loss combines pixelwise photometric differences, flow smoothness, and intensity correlation. Registration accuracy of the proposed and four other registration algorithms is measured by examining tissue integrity, pixelwise

alignment, orientation, and global intensity similarity. The results, tested on two large FLAIRMRI datasets consisting of 700 and 4000 brain volumes, show that the optical-flow registration technique is able to obtain maximal alignment while maintaining structural tissue integrity.