Project Suggestions 2018/19 ____________________________ 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://cmic.cs.ucl.ac.uk/pond to get an idea of other work going on; I am happy to discuss projects in any of those areas too. I am also part of the UCLH Research Hospital initiative, which aims to make hospital data available for machine learning and AI research and development. We have a number of projects available within that activity that you browse here: http://www0.cs.ucl.ac.uk/staff/D.Alexander/Projects/UCLHprojects.18jan19.pdf. Very happy to discuss these. Learning the rules of brain connectivity mapping ================================================ This project will look into learning rules for mapping the connectivity of the human brain from magnetic resonance images. We will use simulations to mimic the process of generating images from known connectivity patterns and use machine learning techniques to emulate the process. The aim is to construct automated techniques that avoid the limitations of current "tractography" algorithms - see for example this recent paper https://www.nature.com/articles/s41467-017-01285-x, which nicely documents the limitations of current technology. Machine Learning for Multi-Contrast Microstructural MRI ======================================================= Applying machine learning techniques to MRI data can yield highly specific insights into small-scale tissue structure. In this project, the student will develop such methods for interpreting in-vivo, multi-contrast MRI scans. Currently we process these scans by calculating a multidimensional spectrum using nonlinear regression (e.g. Slator et al. arXiv preprint 2018, https://arxiv.org/abs/1810.04156). However nonlinear regression has many downsides in this context, such as the need for heavy regularisation and relatively slow model fits. Machine learning has the potential to address some of these concerns. The student will develop machine learning methods which learn the mapping between simulated data and multidimensional spectra. These methods will then be applied to in-vivo MRI data acquired during pregnancy, with the aim of yielding new insights into tissue structure in the fetal brain, lungs and placenta. Skills: python, scikit-learn (or your preferred machine learning toolbox), numpy, matplotlib With Paddy Slator Relating quality of sleep to Alzheimer’s Disease pathology using wearable wrist watch sensors ============================================================================================= The 1946 birth cohort is the oldest continuously studied British birth cohort studies and has followed 5362 individuals since their birth in England, Scotland and Wales during one week in March 1946. Now aged over 70, a small fraction have overt dementia, but estimates suggest that ~1/3 of individuals in this age group may be in the preclinical stages of Alzheimer’s disease. Insight 46 is a study run at UCL, which has recruited 502 members of the 1946 study selected at random from those who attended a clinical visit at 60–64 years. Insight 46 has to date performed detailed brain imaging, including PET and MRI, neuropsychological evaluations and genetic studies on all participants once; and is currently seeing all of them for a follow-up visit. At this follow-visit, activity day is being collected using a Philips Actiwatch Spectrum Plus for a week. To date over 100 individuals have these data collected. In this research project we aim to analyse association between measures of sleep quality, and preclinical Alzheimer’s Disease, and other disease measures such as Parkinsonism. This study provides a unique opportunity to combine these data with such a detailed brain imaging and neuropsychology battery in a birth cohort, and so to investigate the role of sleep impairment in presymptomatic neurodegeneration. We are looking for a student to develop a computational pipeline to extract, preprocess and analyse these data. In particular we would like students to apply skills learnt in their programming module to use Python and open source tools such as Pandas to write a series of scripts capable of: reading actiwatch data files, excluding data containing measurement errors, breaking the data into 24 hour periods and applying literature methods to quantify sleep quality. With Neil Oxtoby, Jon Schott, and Daniele Ravi Image quality transfer ====================== This project uses machine learning to reconstruct brain images of unprecedented quality and resolution. The idea is to learn from very high quality brain images obtained from a unique bespoke MRI scanner and to propagate the information they contain to more every-day data routinely acquired in hospitals. The project builds on an ongoing collaboration with Microsoft Research and uses random forest regression with diffusion tensor imaging. The applications are in neuroscience and in neurological diseases, such as dementia and multiple sclerosis, but the content of the project will focus on developing some specific new machine learning and/or image-analysis techniques that lend themselves specifically to this problem. The TADPOLE challenge: https://tadpole.grand-challenge.org ========================================================== The TADPOLE challenge was set to establish how well we can predict the onset and progression of Alzheimer's disease from standard measurements made from patients (imaging, cognitive tests, fluid markers, etc.). The challenge deadline is passed now and evaluation is under way. However, there are very likely to be future iterations and this project will look into implementing and testing some novel ideas in this prediction challenge using machine learning and image analysis. Machine learning for data harmonisation ======================================= This project will look at the MUSHAC challenge https://projects.iq.harvard.edu/cdmri2018/data, which is about bringing images from different sources into a common space so that they can be analysed together. We will implement and test some deep learning solutions to this problem as well as some basic image analysis techniques to improve the features that go into the learning algorithms. This will place us very well to be ready for the next iteration of the challenge likely to happen next year. Tracking multi-morbidity in the ageing population ================================================= This project will develop methods for identifying groups of patients that follow a similar trajectory of accumulating chronic diseases as they get older. Large collections of health records either from GPs or from hospital admissions can show these trajectories in individuals. The question we address here is what trajectories appear commonly across large groups of individuals. Finding such trajectories can inform on future treatments and care pathways to minimise over prescription of treatments and predict descent into frailty. The project will implement ideas similar to those described in this paper: https://www.nature.com/articles/ncomms5022, and build on them with several methodological enhancements. We will use data sets from the British healthcare system for the first time to reveal the landscape of trajectories. Learning the patterns of clinical phenotypes in Huntington's disease ==================================================================== Huntington's disease (HD) is a devastating neurodegenerative disorder characterised by motor, cognitive and psychiatric symptoms. There are competing hypotheses as to the relation between these symptoms, and which are the most suitable for characterising disease state. In this project we will use clinical data from a worldwide HD study (Enroll-HD) to test whether psychiatric symptoms cluster with motor signs and cognitive features, and hence inform a phenotypic model of HD progression. Contact: Peter Wijeratne Optimising Treatment of Neovascular Age-related Macular Degeneration using Reinforcement Learning ================================================================================================= Background: Age-related macular macular degeneration (AMD) is the leading cause of irreversible blindness in the UK, Europe, and North America. In the UK alone, nearly 200 people develop the blinding forms of AMD every single day. Much of the visual loss in AMD occurs due to the development of choroidal neovascularization (CNV) - so called “wet” or neovascular AMD. In recent years, this condition can be successfully treated with pharmacotherapies that block vascular endothelial growth factor (VEGF). Unfortunately, patients receiving this treatment require intraocular injections on a monthly basis over many years. In order to reduce the burden on patients, while preventing irreversible sight loss, a number of variable treatment regimens have evolved. Reinforcement learning (RL) may afford the opportunity to optimise these treatment regimens - this will be of particular importance as newer agents, potentially longer-acting in some patients, begin to emerge. Moorfields Eye Hospital NHS Foundation Trust (affiliated with UCL Institute of Ophthalmology) has the largest single centre database of eyes treated with anti-VEGF therapy in the world. This consists of 8174 eyes (6664 patients) with 120,756 treatment episodes (figures correct to August 2018).1 In a recent ground-breaking paper, RL has successfully used for optimisation of treatment in patients with sepsis - we propose to adapt this approach for the optimisation of treatment for neovascular AMD.2 Methods: For our model development dataset, we will first collate demographic (e.g., age) and clinical metadata (e.g., visual acuity) for 120,756 treatment episodes. Each of these treatment episodes is associated with at least one retinal optical coherence tomography (OCT) scan (OCT is a form of high resolution imaging that is similar to ultrasound but measurements light reflections rather than sound). We will apply an automated image segmentation algorithm developed as part of the Moorfields-DeepMind collaboration to reduce the dimensionality of these scans.3 This will lead to generation of approximately 100 different structural imaging variables for each image volume (e.g., central retinal thickness). This data will then be coded as multidimensional discrete time series with weekly steps. For each patient, at least 2 years of data will be included, taken from the initial onset of disease. The requirement for an anti-VEGF injection and the suggested subsequent weekly follow-up interval will be taken as the medical treatment of interest. The model will then be trained to optimise patient visual acuity (e.g., to maintain visual acuity >70 letters, the legal limit for driving) and/or anatomic outcome (e.g., a fluid-free retina). A Markov decision process (MDP) - or alternative RL approach - will then be used to model each eye’s environment and trajectory. The RL agent will be deployed to solve the MDP and predict outcomes of treatment strategies. First, we will evaluate the actual treatment of ophthalmologists at Moorfields (need for injection and suggested follow-up interval) and determine the average return of each treatment option. Then, the MDP will be solved using policy iteration, which identifies the treatments that maximize return. A number of models will then be trained on a separate validation dataset before the best model is evaluated on an independent test set. For the latter, patient visual and anatomic outcomes will be analysed when the treatment regimen either corresponded to or differed from that suggested by the best RL model. 1. Fasler, K., Moraes, G., Wagner, S., Kortuem, K. U., Chopra, R., Faes, L., et al. (2018). One and Two Year Visual Outcomes from the Moorfields AMD Database - an Open Science Resource for the Study of Neovascular Age-related Macular Degeneration. bioRxiv, 450411. http://doi.org/10.1101/450411 2. Komorowski, M., Celi, L. A., Badawi, O., Gordon, A. C., & Faisal, A. A. (2018). The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care. Nature Medicine, 353, 1. http://doi.org/10.1038/s41591-018-0213-5 3. De Fauw, J., Ledsam, J. R., Romera-Paredes, B., Nikolov, S., Tomasev, N., Blackwell, S., et al. (2018). Clinically applicable deep learning for diagnosis and referral in retinal disease. Nature Medicine, 31, 1. http://doi.org/10.1038/s41591-018-0107-6 With Pearse Keane at Moorfields Eye Hospital Diagnostic analysis of eye scans using machine learning ======================================================= Acanthemoeba keratitis is a rare but devastating eye condition. In vivo confocal microscopy is a useful tool in providing a detailed analysis of the human cornea and is capable of detecting the cystic organism of acanthamoeba keratitis. The density, location, morophology and distribution of cysts is an important diagnostic and prognostic indicator, however the interpretability of these scans is often difficult and requires highly trained graders to discriminate between acanthemoeba cysts and other features such as white blood cells. This project would involve using machine learning techniques on the Moorfields Eye Hospital database of volumetric confocal image stacks containing a number of eye conditions. The aims are twofold: 1) Diagnosis of acanthemoeba keratitis vs other eye conditions; 2) Quantification of acanthamoeba cysts including cyst detection, segmentation, morphometry, and spatial localisation using machine learning techniques such as the recently published 'U-net for cell-counting, detection and morphometry'1. Correlation of these quantitative parameters with the clinical staging of the disease will be evaluated to identify patients who would need prolonged treatment and to use these characteristics as prognostic indicators for determining disease outcome.​ 1. Falk T, Mai D, Bensch R, Çiçek Ö, Abdulkadir A, Marrakchi Y et al. U-Net: deep learning for cell counting, detection, and morphometry. Nature Methods. 2018;16(1):67-70. With Pearse Keane and Reena Chopra at Moorfields Eye Hospital 3D Ultrasound reconstruction using machine learning =================================================== TBC... Modeling brain microstructure for quantitative neuroimaging =========================================================== Contact: Marco Palombo