Comprehensive Framework for Machine Learning-Enabled Adaptive Optics Scanning Light Ophthalmoscopy

Abstract

Adaptive optics scanning light ophthalmoscopy (AOSLO) enables cellular-resolution imaging of the living human retina by correcting ocular aberrations and isolating light originating from the focal plane. In addition to conventional confocal imaging, nonconfocal detection methods have expanded the capabilities of AOSLO by capturing multiply scattered light that provides complementary structural information about retinal tissue. These approaches have enabled visualization of a wide range of retinal microstructures, including photoreceptor inner segments, retinal vasculature, and immune cells. However, practical limitations remain that restrict the broader use of multichannel nonconfocal AOSLO. Acquisition of multiple detection channels often requires either multiple detectors or sequential measurements, increasing system complexity and acquisition time. Furthermore, the contrast observed in nonconfocal images depends strongly on detector geometry, producing directional contrast that complicates interpretation and prevents consistent representation of cellular morphology across channels. Finally, the small isoplanatic patch of the eye limits the field-of-view (FOV) that can be imaged with diffraction-limited resolution, requiring multiple tiled acquisitions to image larger retinal regions.This dissertation presents computational and hardware-enabled methods to overcome these limitations by enabling rapid multichannel acquisition, unified reconstruction of directional contrast signals, and efficient large FOV retinal imaging. First, a deep compressed multichannel adaptive optics scanning light ophthalmoscopy (DCAOSLO) framework is introduced to accelerate acquisition of nonconfocal imaging channels. Using programmable detection patterns implemented with a digital micromirror device (DMD), multiplexed measurements are acquired and reconstructed using deep learning–based models to recover multiple nonconfocal channels from a reduced set of measurements. This approach significantly reduces acquisition time compared to conventional sequential offset-aperture imaging while preserving channel-specific contrast information. Second, a physics-based forward model is developed to describe the formation of nonconfocal AOSLO images in terms of absorption and phase transfer functions determined by detector geometry. Based on this model, a multichannel fusion algorithm is introduced to integrate information from multiple nonconfocal detection channels into a single unified reconstruction. By jointly incorporating complementary directional contrast information from each channel, the proposed method produces images with improved structural representation and isotropic contrast, enabling more consistent visualization of retinal microstructures. Finally, these computational methods are extended to enable fast large FOV multichannel retinal imaging. By combining accelerated acquisition with multichannel reconstruction and fusion, large retinal mosaics can be efficiently reconstructed while preserving high-resolution structural contrast. This framework enables imaging of retinal areas substantially larger than the isoplanatic patch while maintaining the benefits of multichannel nonconfocal detection. Together, the methods developed in this dissertation establish a computational framework for improving the speed, interpretability, and spatial coverage of adaptive optics retinal imaging. These advances expand the practical capabilities of multichannel AOSLO and provide new tools for studying retinal structure and disease at cellular resolution over clinically relevant spatial scales.