Project Suggestions 2022/23 ____________________________ Here are some projects suggestions for this year. Have a look at my home page http://www.cs.ucl.ac.uk/staff/d.alexander or my groups' websites http://mig.cs.ucl.ac.uk and http://pond.cs.ucl.ac.uk to get an idea of other work going on; I am happy to discuss projects in any of those areas too. Image Quality Transfer with high resolution Connectom data ========================================================== This project extends my group's work on image quality transfer, IQT (Alexander Neuroimage 2017; Tanno Neuroimage 2021; Blumberg MICCAI 2018; Lin MLMIR 2019), which uses machine learning to estimate high quality from low quality images. The project will use very high resolution data sets from the Connectom MRI scanner (Jones Neuroimage 2018), one of four worldwide specialized scanners equipped with super-powered magnetic field gradients that enable unprecedented image resolution. We will start by using existing implementations to demonstrate the potential for enhancing standard neuroimaging data sets by approximating the uniquely high quality Connectom data. Then we will consider alternative implementations of IQT that might exploit the Connectom data better, such as generative adversarial models rather than standard CNNs, and/or explore the additional potential for brain-connectivity mapping that the IQT enhancement provides. Image analysis tools for the Africa brain image archive ======================================================= This project will use an opportunity to collaborate with an emerging network of imaging scientists across Africa, who are collating medical imaging (primarily brain MRI scans) into centralised data bases with the aim of enabling big-data analyses of brain structure and function for comparison with studies made on data from different populations in higher income countries. The data is highly varied, often of low quality, and in a variety of formats. We will work with local groups to identify key tools for harmonizing data sets and perform the first population-wide analyses of data from these regions. Reinforcement learning for tractography ======================================= This project will develop some novel ideas using reinforcement learning to guide brain-connectivity mapping via magnetic resonance imaging and tractography. Tractography has revolutionised our understanding of the connectivity architecture of the brain over the last few decades, as well as given new insight into how brain diseases like Alzheimers and other dementias develop and spread. However, the technique has some substantial limitations in terms of false positive and false negative connections. This project builds on ideas explored in this paper: https://www.sciencedirect.com/science/article/abs/pii/S1361841521001390, and will use reinforcement learning approaches to explore the space of possible strategies for extracting trajectories of brain pathways from diffusion MRI brain-scans. It has the potential to identify radical solutions to some of the major limitations to current ad-hoc approaches. With Ellie Thompson (CMIC) Subtype and stage models of chronic eye disease =============================================== SuStaIn (Young Nature Comms 2018) is an algorithm for discovering disease subtypes from large patient data sets. It defines each subtype by a disease trajectory typically described as a series of events in which biomarkers become abnormal. SuStaIn was originally developed for neurological diesases like Alzheimer’s disease, but adapts naturally for other chronic diseases if sufficient data is available. Collaboration with Moorfield’s eye hospital provides access to appropriate data sets and preprocessing techniques to provide input for SuStaIn. The project will focus on applications to diseases such as age-related maculate degeneration (AMD) and adaptations to the algorithm to make it more suitable for this new application area. MRI Microstructure Imaging through Unsupervised Machine Learning ================================================================ Quantitative MRI and in particular Diffusion MRI provides unique information on tissue microstructure and composition such as axon diameter, cellularity, and blood flow. However traditional approach involve very long processing times requiring non-linear optimisation in every image voxel. We have recently introduced data-driven unsupervised machine learning techniques for this task that provide unbiased parameter maps in a fraction of the time. This enables much more widespread use of the techniques e.g. in cancer assessment and dementia diagnosis. In this project, we will further develop and refine these ideas and work with industry to implement them within their on-board scanner software. More ophthalmology-related projects =================================== https://pontikoslab.com/projects from Nikolas Pontikos and colleagues. Some projects from Ahmed Karim (CMIC) ===================================== Image Quality Transfer using Non-local Sparse Attention and an Application to Low-Field MRI Abstract: Image Quality Transfer (IQT) aims to enhance the contrast and resolution of low-quality medical images, e.g. obtained from low-power devices, with rich information learned from higher quality images. Due to the ill-posed nature of the problem, some sort of regularisation (or prior information from the Bayesian perspective) is required. In this work, we will test different prior models to image quality transfer and evaluate the performance of the approaches using a low-field MRI application. The idea will be to recover contrast enhanced and high-resolved images from low-field ones akin to those obtained using high-field MRI scanners. References: Alexander, Daniel C., et al. "Image quality transfer and applications in diffusion MRI." NeuroImage (2017). Blumberg, Stefano B., et al. "Deeper image quality transfer: Training low-memory neural networks for 3d images." Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Conference. Multispectral Image Segmentation using Machine Learning Abstract: Segmentation algorithms assign a label to every pixel in an image. This has been crucial in various application in medicine, remote sensing, and astronomy. With the increasing spatial resolution of optical systems, different substances tend to have similar spectral signatures, and the traditional image segmentation methods based on shallow and mid-level features have limited space for improvement. This work will investigate a novel machine learning approach for multispectral image segmentation by considering pixel-wise classification of multispectral signals. The approach will be validated using different datasets from different applications in medicine and remote sensing. References: Sivic, Josef, and Andrew Zisserman. "Efficient visual search of videos cast as text retrieval." IEEE transactions on pattern analysis and machine intelligence (2008). Schellenberg, Melanie, et al. "Semantic segmentation of multispectral photoacoustic images using deep learning." Photoacoustics (2022). Quantification of Hebetic Iron Content using MRI Abstract: Iron toxicity is the major cause of tissue damage in patients with iron overload. Iron deposits mainly in the liver, where its concentration closely correlates with whole body iron overload. Different techniques have been proposed for estimating iron content, with liver biopsy being the gold standard despite its invasiveness and influence by sampling error. Recently, magnetic resonance imaging (MRI) has been established as an effective technique for evaluating iron overload by measuring T2* in the liver. However, various factors associated with the adopted analysis technique, mainly the exponential fitting model and signal averaging method, affect the resulting measurements. In this work, we will implement a fast algorithm to estimate T2* constant in MRI that can be subsequently used for quantification of iron overload in the liver. References: 1. Ahmed K. Eldaly et al. (2023) ‘’A General Framework for Hepatic Iron Overload Quantification using MRI’’ Elsevier Journal of Digital Signal Processing (to appear). 1. Elsayed Ibrahim et al., (2016). Influence of the analysis technique on estimating hepatic iron content using MRI. Journal of Magnetic Resonance Imaging. 2. Manya Afonso, (2010). An augmented Lagrangian approach to the constrained optimization formulation of imaging inverse problems. IEEE Transactions on Image Processing. An Unsupervised Learning Approach for Bacterial Detection in Optical Endomicroscopy Images Abstract: Pneumonia is a major cause of morbidity and mortality of patients in intensive care. Rapid determination of the presence and gram status of the pathogenic bacteria in the distal lung may enable a more tailored treatment regime. Optical Endomicroscopy (OEM) is an emerging medical imaging platform with preclinical and clinical utility. Pulmonary OEM via multi-core fibre bundles has the potential to provide in vivo, in situ , fluorescent molecular signatures of the causes of infection and inflammation. This work will investigate the performance of an unsupervised approach for bacterial detection in datasets of optical endomicroscopy (OEM) lung images. This algorithm will be validated by simulations conducted using synthetic datasets. Analysis is then conducted using ex vivo lung datasets in which fluorescently labelled bacteria are present in the distal lung. References: A. Eldaly, et al. “Bayesian bacterial detection using irregularly sampled optical endomicroscopy images” Elsevier Medical Image Analysis, 2019. A. Adler, et al. ’’Sparse coding with anomaly detection’’ Journal of Signal Processing Systems, 2015.