A Neural Network Approach to Quantify Blood Flow from Retinal OCT Intensity Time-Series Measurements

Abstract

Many diseases of the eye are associated with alterations in the retinal vasculature that are possibly preceded by undetected changes in blood flow. In this work, a robust blood flow quantification framework is presented based on optical coherence tomography (OCT) angiography imaging and deep learning. The analysis used a forward signal model to simulate OCT blood flow data for training of a neural network (NN). The NN was combined with pre-and post-processing steps to create an analysis framework for measuring flow rates from individual blood vessels. The framework's accuracy was validated using both blood flow phantoms and human subject imaging, and across flow speed, vessel angle, hematocrit levels, and signal-to-noise ratio. The reported flow rate of the calibrated NN framework was measured to be largely independent of vessel angle, hematocrit levels, and measurement signal-to-noise ratio. In vivo retinal flow rate measurements were self-consistent across vascular branch points, and approximately followed a predicted power-law dependence on the vessel diameter. The presented OCT-based NN flow rate estimation framework addresses the need for a robust, deployable, and label-free quantitative retinal blood flow mapping technique. Many retinal diseases are associated with abnormalities in perfusion with primary examples including age-related macular degeneration, diabetic retinopathy, and glaucoma 1-6. Much of our current understanding in this area is derived from fluorescence angiography (FA) and, more recently, optical coherence tomography (OCT) angiography. These tools provide the morphology of the retinal vasculature, e.g., vessel diameter and capillary drop-out, but do not quantify retinal blood flow directly. OCT imaging is commonplace in ophthalmology and an OCT-based flow imaging technique could therefore be rapidly adopted in research and clinical settings. This is especially true if the technique can be compatible with the design of commercial OCT platforms. Traditionally, flow measurements in OCT have been based on Doppler (phase-based) techniques, which measure the axial component of the blood flow 7-13. The primary barrier to the use of Doppler OCT in the retina arises from the need to calculate total flow from axial flow using a scale factor that is inversely related to the cosine of the angle between the flow vector and the imaging beam (i.e. Doppler angle α). With α near 90°, a small error in the measurement of α leads to large errors in the measured total flow 14. This has limited Doppler techniques to a small region near the optic nerve head where vessels are oriented to avoid α = 90°1 5. Across most of the retina, where vessels are oriented with α ~90°, only techniques such as multi-beam Doppler OCT have been successful in accurately measuring flow 16-19. However, the use of multiple, non-colinear beams in three-dimensional imaging significantly increases the hardware complexity. The required modifications to microscope and OCT hardware are barriers in this case to a broader adoption of the tech...

Deep Learning in Biomedical Optics

“…While normally requiring specific imaging protocols, the raw OCT measurements contain structural features that may be recognized by a CNN. Reports have shown that angiographic image and blood flows can be predicted by mere structural image input without specific OCTA or DOCT protocols [190–192]. For example, Braaf et al [190] showed that DL enabled accurate quantification of blood flow from OCT intensity time‐series measurements, and was robust to vessel angle, hematocrit levels, and measurement SNR.…”

“…Reports have shown that angiographic image and blood flows can be predicted by mere structural image input without specific OCTA or DOCT protocols [190–192]. For example, Braaf et al [190] showed that DL enabled accurate quantification of blood flow from OCT intensity time‐series measurements, and was robust to vessel angle, hematocrit levels, and measurement SNR. This is appealing for generating not only anatomical features, but also functional readouts using the simplest OCT imaging protocols by any regular OCT device (Fig.…”

“…( a ) Examples of using DL to predict blood flow based on structural OCT image features (reprinted from [190]). ( b ) Example of deep spectral learning for label‐free oximetry in visible light OCT (reprinted from [193]).…”

Oxycytosis and the role of triboelectricity and oxidation in bacteria clearing from the bloodstream

“…The blood flow in the vessels is fast, and all cells move quickly [ 41 ]. Not only catching pathogens but also their killing should be fast and effective [ 42 ].…”

Using the Dynamic Forward Scattering Signal for Optical Coherence Tomography based Blood Flow Quantification

“…All data in this work were acquired using a swept-source OCT system (center wavelength 1060 nm, 100kHz A-line rate), and the OCT system, flow phantom, and beam scan protocols are described in greater detail in Refs. [10,11]. dependence on Doppler angle is seen, but it is likely that this is a measurement artifact rather than an intrinsic response.…”