1st Supervisor: David Firmin, Imperial College London
2nd Supervisor: Sonia Nielles-Vallespin, Imperial College London
Additional Supervisors: Guang Yang, Andrew Scott, Pedro Ferreira, Imperial College London
Clinical Champion: Zohya Khalique, Imperial College London
Aim of the PhD Project:
- Reduce the number of repetitions needed to post-process in vivo DT-CMR towards a single-breath-hold, with a model-free tensor prediction by convolutional neural networks.
- Integrate these trained models in the scanner’s on-line post-processing pipeline.
Project Description / Background:
Diffusion tensor cardiac MR (DT-CMR) is a unique and emerging technique capable of providing information on myocardial microstructure in-vivo. DT-CMR quantifies water diffusion in the myocardium, which is constrained by the myocardial micro-architecture. A diffusion tensor is calculated for each voxel and many different diffusion parameters can be extracted from the tensors including tensor orientation measures which have been shown to relate to the orientation of the local cardiomyocytes and their sheetlet structure [Ferreira 2014, Nielles-Vallespin 2017]. This technology provides an insight into the poorly understood link between cellular contraction and macroscopic cardiac function. It offers huge potential for novel microstructural and functional assessment, and for the development and evaluation of novel therapeutic approaches. However, considerable technical challenges exist to translate this technology to clinical routine.
In vivo DT-CMR requires the rapid acquisition of multiple diffusion weighted images with diffusion encoded in at least 6 different directions in 3D space. Rapid acquisition translates to low SNR images, and therefore multiple repetitions are commonly acquired to build up SNR. Each repetition requires an extra breath-hold for the patient, with current clinical research protocols requiring approximately 10 breath-holds for every single DT-CMR slice at a single time point in the cardiac cycle. Typically, the series of signal intensities for each voxel is then fitted to a rank-2 diffusion tensor by using a linear least-square model.
Recently the work by Alliota et al. showed that it is possible to train a convolutional neural network (CNN) to reconstruct tensor measures from reduced subsets of DTI brain magnitude signals, including using only 3 diffusion encoded directions, which is less than the minimum needed to fit a tensor with a conventional least-square approach [Alliota 2020]. This model-free approach allows therefore for shorter acquisitions.
In this project we propose building on the work from Alliota et al. by reconstructing diffusion tensors for in vivo cardiac DT-CMR with a model-free deep learning approach with the aim of reducing the number of repetitions, and therefore breath-holds needed to only one. We will train different convolutional neural network designs and instead of training the reconstruction of a fixed set of tensor parameter results, we instead propose to reconstruct the full diffusion tensor elements which will then allow us to extract all the common tensor orientation and shape parameters without any constraint.
The CMR unit at the Royal Brompton Hospital has a team of physicists and clinicians dedicated to in vivo DT-CMR development since 2009. We have acquired a considerable amount of high quality scans in both healthy volunteers (>350 scans) and patients (>200 scans) (STEAM EPI sequence with a monopolar diffusion encoding scheme (b-values=[150,750])). All these scans have been carefully post-processed by a team of experienced users. We will use these data to develop and train our convolutional neural networks.
The candidate will have an academic background in Physics, Mathematics, or Computer Science with a strong interest in AI and a love for programming.
Figure 1: Four different DT-CMR parameter maps for one healthy scan. Left: conventional least-square tensor calculation using 11 breath-holds. Centre and right: U-Net tensor reconstruction using a reduced number of images acquired over 5 and one breath-hold.
- Ferreira PF, J Cardiovasc Magn Reson. 2014; 16: 87.
- Nielles-Vallespin S, J Am Coll Cardiol. 2017; 69: 661-676.
- Aliotta E, Nourzadeh H, Patel SH. Extracting diffusion tensor fractional anisotropy and mean diffusivity from 3-direction DWI scans using deep learning. Magn Reson Med. 2020.