A-EYE: Integrating In-silico Modelling and Deep Learning to Optimize Diagnosis and Treatment of Wet Age-related Macular Degeneration

点击翻译按钮获取中文摘要

Age-related macular degeneration (AMD) is one of the leading causes of blindness. In its 'wet' form, abnormal new vessels grow in the retina, causing bleeding, scarring and physical deformation of the retinal structure. The macula being the part of the eye responsible for high resolution, central vision, this disruption of the retinal structure is particularly detrimental to vision. Currently, treatment for wet AMD consists of intravitreal injections of a molecule inhibiting growth of the neovasculature. Regardless of how much of a revolution this kind of therapy has been for conserving sight, it does not address the underlying cause of wet AMD and therefore injections must be repeated often, representing a substantial burden for the patient and cost for the health care system. Furthermore, patients respond very differently to the treatment.Optical Coherence Tomography Angiography (OCTA) is a novel, quick, non-invasive, and high-resolution tool for imaging the retinal vasculature. OCTA allows novel insight into the retinal microvasculature in healthy and diseased eyes. It is becoming increasingly used as a diagnostic tool for wet AMD. One of the applications of OCTA is to classify neovasculatures into subtypes which have different response to the current treatment. Additionally, OCTA provides a three-dimensional representation of the retina, its vasculature, and the blood flow therein. The three-dimensional information has been shown to provide more efficient biomarkers (e.g., abnormal fluid volumes) than the previous two-dimensional information (e.g., central retinal thickness) [1].Despite its extensive use in medical imaging, deep learning methods and in particular convolutional neuron networks (CNN) have yet to be applied to classifying neovasculatures from OCTA scans and is one of the goals of this PhD project. Additional novelty will be in the use of three-dimensional segmentations of the microvasculature as an input for the algorithm. Furthermore, the learning algorithm will be informed by computed parameters of the vasculature from our in-silico models to potentially improve its accuracy and explainability, two of the major challenges inherent to artificial intelligence. Currently, most research has been focussing on learning from data, with little work on computer simulations (I.e., in-silico models) of the physiology. Such models rely on physiologically informed assumptions and mathematical equations to provide understanding of the, possibly nonlinear, relationships at work within a system [2]. Furthermore, they allow for quick, inexpensive experiments 'in-silico' which are an efficient tool for trials of new treatment protocols. The function of retinal blood flow in the retina in both healthy and diseased eyes will be investigated in-silico. Later, blood flow and metabolic response will be combined in models of the diseased eye to investigate patient's responsiveness to treatment and explore the possibilities for new treatment regime and new therapy modalities. As part of this, a pipeline (see Fig. 1) will be developed to go from OCTA images, to segmented vasculature, to personalised in-silico models of blood flow and drug treatment [3].Clinical data will be collected from the St Paul's Eye Unit work at the Royal Liverpool University Hospital. The close collaboration with the ophthalmic team is essential for the development of accurate models as well as maintaining a degree of translatability into the clinic.