Dynamic retinal blood flow analysis: Heartbeat-correlated artery and vein identification in laser speckle flowgraphy

Laser speckle flowgraphy (LSFG) is a non-invasive technique for evaluating retinal and choroidal blood flow, estimating flow velocity using the mean blur rate (MBR) from speckle pattern changes. An LSFG video consists of 120 time frames over approximately four heartbeats. The traditional analysis tool, total retinal artery and vein analysis (TRAVA), requires manual identification of the optic disc and setting up regions of interest (i.e., placing an annulus grid). This manual process is often tedious and subjective. In this study, we introduce an optimized deep-learning framework (utilizing nnU-Net, version 2.1) using temporal blood flow variations across heartbeats as multichannel inputs to automatically identify retinal arteries and veins in LSFG. For model training, an ophthalmologist manually traced arteries and veins in color fundus photographs, which were then converted into LSFG labels by an LSFG expert. Data from 40 patients with both LSFG and fundus photographs were used: 30 for training and ten for testing. The proposed model achieved Dice coefficients of 0.78 ± 0.06 for arteries and 0.79 ± 0.06 for veins, significantly outperforming the state-of-the-art AutoMorph’s core approach (retrained by the same data), which had Dice coefficients of 0.61 ± 0.16 and 0.63 ± 0.07, respectively (p-values ⪅ 0.01). Our method also yielded significantly higher recall and F1 scores in identifying artery and vein segments at the optic disc compared to TRAVA (p-values ⪅ 0.01). For vessel segment counting and blood flow quantification at the optic disc annulus grid, the proposed method showed no significant differences from manual tracing (p-values ≫ 0.05). Overall, nnU-Net with normalized LSFG time frames as multi-channel inputs can automatically identify arteries and veins without manual intervention. This automated method enables the use of LSFG for larger datasets, potentially increasing its adoption and utility.