Host: Professor Feng Liu
Speaker: Yu He
Abstract:
Zoom link: https://uqz.zoom.us/j/88012244836
Magnetic resonance imaging (MRI) is a medical imaging device using electromagnetic fields to reveal the human body's anatomical information. High spatial resolution images can provide excellent soft-tissue contrast for accurate diagnosis and treatment. However, high-resolution images cannot always be obtained owing to the acquisition limitations such as insufficient scanning time and moving human organs. As a result, the in-plane/in-slice resolution of MRI images is usually clinically confined to several millimeter ranges. Such pixel/ voxel resolution may not provide sufficient pathology depiction for precise diagnosis. One promising approach, the super-resolution (SR) technique, can enhance MRI spatial resolution effectively and efficiently.
SR techniques can recover detailed contours and textures of images, thus improving the image quality for various applications. It has been extensively used in image decryption, video surveillance, and medical imaging. Recently, SR techniques combined with deep learning (DL) have significantly improved MR image reconstruction. Convolutional neural networks (CNNs), one type of DL, show their superior performance in image quality and operation time compared to other conventional SR approaches. However, since their inadequate representation of reconstructing high-frequency information, current CNN-based methods still lack the capability of recovering detailed information, leading to blurring images. Moreover, DL-based SR techniques employ data-driven optimization procedures, which heavily rely on data quantity. Clinically, collecting sufficient image data becomes challenging because of privacy limitation, cross-vendor, and cross-scanner variation. As a result, insufficient image data causes severe overfitting, degrading CNN performance.
The thesis provides three principal contradictions: Firstly, a frequency-based network, the channel splitting network (DCSN), is proposed for 2D MRI super-resolution. It aims to split the frequency bands of MR images and carry their high-frequency components into deeper cascades of the network, enabling a better representation of MRI. Secondly, an affine transformation data augmentation, optical flow-based registration approach, is adopted to increase the data volume. The method allows CNNs to be trained with few data volumes. Lastly, a zero-shot, unsupervised scheme is proposed to solve the problem of insufficient data further. Ideally, the proposed method can allow lightweight CNN to be trained using only a single low resolution (LR) image.
Biography:
Yu He is currently an MPhil student under the supervision of Professor Feng Liu and Dr Fangfang Tang. His main research interest includes MRI methods development, super-resolution, data augmentation, and deep learning.