OptiFlow: Deep Learning-Based Motion Estimation and Frame Interpolation for Brain MRI

Motivation: High-quality MRI acquisitions often require long scan times causing motion artifacts and image distortions. Accelerated scan times are essential to improve efficiency and reduce image distortions. Goal(s): The goal of this work is to interpolate intermediate unacquired slices from an accelerated acquisition of acquired slices. Approach: Our proposed method was pretrained on video frames, fine-tuned on 221 randomly selected healthy subjects from the OASIS-1 Dataset, and tested on a cohort of 10 subjects. Results: The proposed model demonstrates the ability to interpolate intermediate MRI slices with relatively high similarity to the ground-truth. Impact: This work serves as a proof of concept, in which our optical flow-based deep learning technique has the potential to reduce MRI acquisition times, overcome issues of image distortions, and enhance overall image quality.