Student: Gavin Seegoolam
The aim of the project is to develop machine learning approaches for the reconstruction, synthesis and analysis of MR images. In particular, we aim to develop approaches in which image acquisition, reconstruction and analysis are not carried out in a sequential fashion, but are integrated into a joint reconstruction and analysis framework. This will enable feedback between the image acquisition, reconstruction and analysis stages, leading to improved reconstruction as well as analysis.
The project will address the following challenges:
- How can we learn image representations that can serve as effective and informative priors for image acquisition/reconstruction?
- How can we learn image representations that can be used for image synthesis/analysis tasks, e.g. image super-resolution or image segmentation?
- How can these image representations be combined for joint reconstruction and analysis?
Magnetic Resonance Imaging (MRI) is an indispensable tool diagnosis and interventions. Yet, MRI suffers from a relatively slow data acquisition process since the acquisition speed is constrained by the serial fashion in which the k-space data, necessary for reconstructing the images, is acquired. A common approach for accelerating the MRI acquisition is compressed sensing which enables the reconstruction of images from undersampled k-space measurements. By exploiting sparsity, compressed sensing turns the ill-posed reconstruction problem into a mathematical optimization problem that is both sound and computationally tractable. Machine learning techniques, such as dictionary learning, has also been demonstrated as an effective method to further improve the reconstructions from highly undersampled measurements [Caballero 2014].
Recently, in the field of computer vision and machine learning, deep learning techniques have emerged as very effective machineries to solve a variety of imaging problems. In particular, convolutional neural network (CNN) achieve state-of-the-art performance for a variety of imaging problems, such as classification, segmentation and object detection. Surprisingly, however, CNNs have not yet been applied to compressed sensing or image reconstruction problems.
In this project, we explore the use of different deep learning approaches for MRI reconstruction and segmentation from undersampled k-space measurements. This will lead to application-specific MRI reconstruction algorithms that enable the joint optimisation of reconstruction and segmentation, enabling the user balance reconstruction fidelity and segmentation/registration performance depending on the clinical application. Our initial approach to realize this joint reconstruction and segmentation will be based on CNNs but later we will explore the use of other deep learning approaches including adversarial networks.
The approaches developed in the PhD project will be tested on fetal and cardiac MR images for which fast image acquisition due to motion/movement is important.