Computational Pipeline for Retinal Capillary Blood Flow Measurement with Adaptive Optics Line Conf

Overview of the AOLCO Blood-Flow Pipeline

This work describes a computational pipeline designed to measure erythrocyte velocity in retinal capillaries from adaptive optics line confocal ophthalmoscopy (AOLCO) video data. The pipeline is provided as an interactive 3D Slicer extension and as a batch command-line tool that share a single analysis core. It stabilizes AOLCO capillary videos, suppresses stationary structures, and returns blood-flow velocities in physical units suitable for capillary-level analysis.

Preprocessing: Illumination Correction and Closed-Eye Frame Detection

The first stage of the pipeline performs basic video preprocessing. Illumination nonuniformities are corrected to improve frame-to-frame consistency. The preprocessing stage also identifies and excludes closed-eye frames, reducing the inclusion of frames corrupted by blinks or other transient artifacts prior to registration and velocimetry.

Registration: GPU Translation Initialization, Affine and B‑Spline Alignment

Registration is a critical stage because eye motion in AOLCO acquisitions can produce linear and nonlinear distortions with displacements on the order of tens to hundreds of pixels between frames. The pipeline uses a multi-step registration strategy:

• A fast, robust GPU-based translation-registration algorithm provides initialization and can accommodate large eye movements (e.g., microsaccades). This step is intended to cope with substantial inter-frame displacements before finer alignment.
• Following translation initialization, an affine registration aligns frames to a reference frame.
• When necessary, B-Spline registration is applied to correct residual nonlinear deformations.

The authors report that the GPU translation algorithm accelerates alignment substantially versus a baseline method, running approximately 3.6× faster at 100 Hz and up to 38× faster at 400 Hz, enabling efficient processing of high-frame-rate AOLCO datasets.

Postprocessing: Frame Differencing and Capillary Network Projection

After registration, adjacent registered frames are differenced to suppress stationary structures (for example, tissue background and static vasculature). From the frame-difference video, a 2D projection is generated that highlights the capillary network. This projection is used as a visualization and navigation image for identifying capillary segments for subsequent velocimetry.

Velocity Measurement: Segment Selection, Spatiotemporal Images, and Radon Transform

Velocity measurement proceeds as follows:

• An operator manually identifies capillary segments on the 2D projection image generated in postprocessing. The pipeline relies on manual selection of segments for which spatiotemporal analysis is desired.
• For each selected segment, the pipeline generates a spatiotemporal image by sampling the frame-difference video along the segment over time. These images encode the motion of erythrocytes as slanted traces in space–time coordinates.
• Blood-flow velocity curves are obtained from the spatiotemporal images using the Radon transform, which extracts the slope of erythrocyte traces and converts that slope into a physical velocity measurement.

Validation and Performance Metrics

The pipeline’s velocimetry was validated using synthetic ground-truth flow data reported by the authors. In those validation experiments the velocimetry recovered speed within an error of 4% (reported value).

Performance and quality metrics reported across the full cohort include:

• Dataset size: 2,344 videos from 57 subjects, with acquisition frame rates ranging from 30 Hz to 400 Hz (reported dataset summary).
• The registration approach improved frame structural similarity (SSIM) on every acquisition in the cohort, increasing mean SSIM from 0.73 to 0.86 (reported improvement).
• Reported typical eye motion magnitude handled by the method was approximately 90 μm per frame on average.
• The GPU translation-registration algorithm achieved substantial speedups relative to the baseline, with reported acceleration ranging from about 3.6× (at 100 Hz) to 38× (at 400 Hz).

The authors state that the recovered blood flow velocities are physiologically plausible and of the same order as previously published adaptive-optics measurements.

Software Delivery and Practical Use

The analysis framework is delivered in two user-facing forms: an interactive extension for 3D Slicer and a batch command-line tool. Both interfaces use the same analysis core. The pipeline workflows—preprocessing, registration, postprocessing, and Radon-transform–based velocimetry—are implemented to support both interactive review and automated processing of large AOLCO datasets.

Limitations, Preprint Status, and Data Availability

This work is presented as a preprint and has not been certified by peer review (source statement). The abstract supplies summary-level results and key performance metrics, but the source abstract does not include full implementation details, parameter lists, user interface screenshots, or direct download links for the software. For additional technical details, code availability, and full experimental protocols, consult the full preprint PDF and any repository links provided in the manuscript.

The authors declared no competing interests and reported NIH funding (1OT2OD038131) in the source article.