UC San Diego

Age-related Macular Degeneration (AMD) is a degenerative eye disease that leads to central vision loss, making early detection of AMD pathology, especially choroidal neovascu- larization (CNV), crucial for effective treatment. Optical Coherence Tomography Angiography (OCTA) emerges as a promising tool for AMD diagnosis by visualizing pathogenic vessels in the subretinal space. However, OCTA’s clinical utility is hindered by its early developmental stage. In this thesis, we explore OCTA’s potential in AMD diagnosis from perspective view of deep learning-based classifiers, exploiting its capacity to discern hidden patterns and process 3D data efficiently. To our knowledge, this is the first study exclusively utilizing OCTA for AMD stage grading.Our research uncovers significant challenges in developing deep learning-based AMD stage graders, including the scarcity of suitable OCTA datasets and variability in image quality due to layer segmentation errors during OCTA generation. To address these challenges, we curate a substantial dataset comprising thousands of OCTA samples with clinical annotations. We develop a 2D classifier based on OCTA projections and assess the impact of segmentation errors qualitatively and quantitatively, proposing two potential solutions. Firstly, we advocate for direct analysis of 3D OCTA volumes using a 2D convolutional neural network trained with additional projection supervision. Secondly, in the absence of 3D data, we propose leveraging cross-instrument OCTA data and employing style transfer techniques, such as CycleGAN, to facilitate domain conversion. To enhance CycleGAN’s performance for classification tasks, we integrate class-related constraints during training. Experimental results demonstrate over 80% accuracy on a four-stage grading task, surpass- ing human expert accuracy of 60%, irrespective of segmentation errors. Our findings underscore OCTA’s potential as a promising modality for differentiating AMD disease stages and highlight the significance of our contributions in advancing OCTA-based AMD diagnosis.