A Deep Learning Model for Susceptibility Artifact Correction in Echo Planar Imaging
by Abdallah Zaid Alkilani, M.S. in Electrical and Electronics Engineering
Advisor: Assoc. Prof. Emine Ülkü Sarıtaş
Co-advisor: Assoc. Prof. Tolga Çukur
Date: Thursday, September 08, 2022 at 11 am
Join Zoom Meeting
Meeting ID: 251 006 7074
Diffusion Weighted Imaging (DWI) is a Magnetic Resonance Imaging (MRI) technique that probes the Brownian motion of water molecules within biological tissue, in vivo and non-invasively. The most commonly employed sequence for DWI is Echo Planar Imaging (EPI), where the entirety of $k$-space is collected in a zigzag manner in one shot after a single diffusion preparation step. EPI is preferred due to its robustness to motion, and it meets the high signal-to-noise ratio efficiency and short acquisition duration demands of DWI. However, EPI suffers from severe susceptibility-induced artifacts that arise due to strong readout gradients and low bandwidth along the phase-encode (PE) direction. These artifacts are caused by magnetic susceptibility differences and manifest as geometric and intensity distortions. Postprocessing methods are extensively used to correct for these artifacts, particularly reversed PE techniques that utilize images acquired in reversed PE directions to deduce the susceptibility-induced displacement field. While many non-learning methods exist for the reversed PE approach, they are relatively time consuming and require instance-specific optimization. Only a few recent works have explored the benefits of employing deep learning to speed up the reversed PE approach. These methods rely on unwarping correction with a predicted displacement field that maximizes image similarity.
This thesis proposes a deep unsupervised Forward-Distortion Network (FD-Net) for correcting susceptibility artifacts. FD-Net speeds up the correction while explicitly constraining measurement fidelity for enhanced correction performance. This technique employs an encoder-decoder architecture to predict the field as well as the corrected image from the input reversed PE images. Using the field to forward-distort the predicted image in both PE directions should explain the input reversed PE images, thereby enforcing consistency to input data. This forward-distortion approach relies on matrix operations and is computationally efficient. Two different multiresolution strategies are considered: a multiscale strategy where earlier stages of the decoder produce lower resolution field and image predictions, and a multiblur strategy where the full resolution predictions are progressively blurred. Both strategies aim to boost performance by enforcing consistency across different scales and blurs, effectively speeding up convergence and circumventing local minima. In this thesis, variations of the multiresolution strategies are considered and the highest performing strategy in terms of quantitative image quality metrics is chosen.
The performance of FD-Net is evaluated in comparison to two recent deep learning methods from the literature and a supervised baseline method based on FD-Net. A classical unwarping-based method is used as the gold standard reference. Extensive slice-wise, subject-wise, visual, and quantitative image quality assessments are performed. The results demonstrate that FD-Net surpasses the comparison deep learning methods, and outperforms the supervised baseline in terms of predicted image quality while maintaining robust field predictions. Hence, the forward-distortion model presents a better-conditioned problem for distortion correction when compared to unwarping-based approaches. This thesis concludes that FD-Net provides a novel paradigm for the susceptibility artifact correction problem that better constrains fidelity to the measurement data.