Predictive Modeling of Age-related Macular Degeneration with Multimodal and Interpretable Convolutional Neural Networks
Principal Investigator:
Dr Murat Ayhan
Approved Research ID:
36940
Approval date:
July 23rd 2018
Lay summary
Age-related macular degeneration (AMD) is a major cause of blindness among elderly citizens of the developed world. It is characterized by the degeneration of the central part of the retina and typically develops from an early form with little vision impairment to one of two late forms - dry or wet AMD. The risk factors include age, race, genetic heritage, and lifestyle. Unfortunately, AMD treatment is not well-established. The recent studies on AMD pathogenesis indicate a significant overlap between the underlying mechanisms of the seemingly distinct forms. It is essential to investigate this overlap and the underlying processes for the development of AMD treatments and to find biomarkers that allow to predict disease progression. Recently, deep learning algorithms have revolutionized computer vision and image analysis. In particular, convolutional neural networks (CNNs) have emerged as powerful image analysis and prediction tools. They have surpassed human-level performance on challenging medical tasks like skin cancer classification from dermoscopic images and arrhythmia detection from ECG signals. In our previous work, we used CNNs for diabetic retinopathy (DR) detection from retinal images. Our CNNs for DR detection surpassed the recommendations of the British Diabetic Association and the thresholds set by the NHS Diabetes Eye Screening program. Now, as a team with joint expertise in computer science, bioinformatics, neuroscience and ophthalmology, we want to study AMD with CNNs for patient stratification and develop methods to explain CNNs' decisions in a medical setting. We will use the UK Biobank data to develop multimodal CNNs that will exploit the statistical regularities from both retinal photography and optical coherence tomography (OCT) simultaneously. Since each imaging technique captures different aspects of the underlying disease pathology, we expect to obtain novel biomarkers that will help us explain the AMD mechanisms as well as discern the stages and subtypes of the disease with high accuracy and confidence. We will preprocess and analyze the data on our computing clusters located at our institute's facilities. We will also discuss our findings with ophthalmologists and evaluate them in a clinically relevant manner. We will conclude the study in two years from the receipt of access to data. Our work addresses the UK Biobank's stated purpose by improving the early diagnosis of AMD and promoting health among elderly citizens.