Zamoun, Inès; (2025) Developing an Automated AI Pipeline for Blood Flow Extraction in Retinal Vessels Imaged with Adaptive Optics Scanning Laser Ophthalmoscopy (AOSLO). Masters thesis (M.Phil), UCL (University College London).

Text Zamoun_10211970_thesis.pdf Access restricted to UCL open access staff until 1 September 2026. Download (14MB)

Abstract

The eye, often described as a window to the brain, shares many anatomical and pathological similarities with it. Leveraging the optical clarity of the lens and vitreous body, our work provides access to neurological and vascular structures within the retina, positioning ophthalmology at the forefront of imaging and analytical technologies. One of the most impactful advancements is Adaptive Optics Scanning Laser Ophthalmoscopy (AOSLO), which enables cellular-resolution visualization of the retinal vasculature. In this thesis, we focus on the structure and function of large retinal vessels. We present a fully automated pipeline that tracks vessels across AOSLO XT video frames and extracts blood flow metrics by integrating computer vision techniques and deep learning. Our object detection stage uses YOLO (You Only Look Once), enhanced with pre- and post-processing and custom contour-based logic, achieving 93% mean average precision (mAP) and 85% tracking accuracy within 8ms and 20ms per image, respectively. This represents a significant improvement over manual methods that take over 11 seconds per detection. To estimate blood flow, our system computes red blood cell (RBC) velocity by extracting the orientation of streaks in XT images through angle-based analysis (Radon, FFT, Hough Transform) and compensates for vessel inclination using α correction. Diameters are estimated via Gaussian intensity profiling. This yields velocity measurements from 3.2 mm/s to 53 mm/s for vessel diameters between 23 μm and 150 μm, consistent with Palochak et al. (2013). Our pipeline processes each frame in 0.029 seconds, offering a 1070× speedup over manual and a 355× improvement over semi-automated methods (Houston et al., 2021). Future directions include adapting this pipeline for smaller vessels using either YOLO or U-Net. This work lays the foundation for scalable and reproducible analysis of retinal blood flow and contributes to early detection tools for neurovascular conditions such as multiple sclerosis and diabetic retinopathy. Keywords: AOSLO; Retina Blood Flow; Computer Vision; Deep Learning; Vascular Imaging.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item