基于GPU高速并行计算实现逆信噪比-复值退相关OCTA实时成像

Inverse SNR and Complex-Valued Decorrelation OCTA Real-Time Imaging Based on GPU High-Speed Parallel Computing

在光学相干层析血流造影(OCTA)系统的实际应用中,高质量数据的采集受到多种因素的干扰,如屈光调节、扫描区域移动、动态成像过程中受试对象眼睛状态波动等。笔者构建了一种基于图像处理单元(GPU)的OCTA数据实时处理框架,使用C++和CUDA开发系统软件,实现了逆信噪比-复值退相关光学相干层析血流造影(ID-OCTA)的实时信号处理与图像显示,线处理速度达到了365 kHz。同时,通过闪烁光刺激诱发小鼠视网膜功能性充血实验,证明了本研究实现的OCTA投影图像实时显示功能有助于操作人员调节系统,监测受试对象的状态,从而提高数据采集成功率。

Objective Currently,most commercial optical coherence angiography(OCTA)systems lack a real-time display of en face OCTA images,which makes it difficult for operators to obtain intuitive feedback on data quality and adjust the system quickly and accurately in a single acquisition of OCTA volume data.In the process of dynamic acquisition of OCTA volume data,determining the state changes of the subjects is difficult,resulting in invalid data acquisition.In an experiment on flicker light-induced functional retinal hyperemia,which provides a new perspective for the early screening of human diabetic retinopathy,the continuous collection of multiple groups of three-dimensional data may be invalid because of the poor quality of one group,thereby wasting data processing time.Therefore,a real-time display of the experimental results is required.Although GPU-based OCTA data real-time processing methods have been proposed,the speed of the existing real-time processing methods still needs to be improved to adapt to high-speed scanningOcTAsystems.Methods It is developed on a spectral-domain OCT(SD-OCT)system.Limited by the frame grabber,the maximum acquisition line speed of the system was 120 kHz in the high-bit-depth mode and 250 kHz in the low-bit-depth mode.An optical coherence angiography algorithm based on the inverse signal-to-noise ratio(SNR)and complex-valued decorrelation(ID-OCTA)was used to extract blood signals by adaptive SNR and achieve high-quality angiography.The sum of absolute differences(SAD)algorithm was used to register OCT images,and the retinal OCT images were segmented by a vertical gradient distribution,which is convenient for fast parallel processing on a Graphics Processing Unit(GPU).This study proposes a real-time processing framework based on a GPU(Fig.1),which uses texture memory to realize fast interpolation and filtering calculations and the CUDA stream to mask the time delay of data transmission between the host and GPU.We developed a real-time processing program using C++and CUDA and a multithread system control program using the C++and MFC libraries.To compare the guiding effect of the real-time data processing method in this study and the method using only a CPU,two real-time display modes were used for data acquisition:en face OCTA images and cross-sectional OCT images.Moderately experienced operators collected multiple groups of data in these modes within 4O s.Three sets of data were collected continuously in 12 s to simulate the dynamic acquisition of OCTA volume data.The quality of the collected data was evaluated using the en face OCTA image quality index.In the flicker light-induced functional retinal hyperemia experiment in mice,the experimental success criteria and quantification parameters were set.Operators conducted multiple experiments to compare the experimental success rates of the two real-time display modes.Results and Discussions The en face OCTA image real-time display was realized in the system with a 250 kHz line scanning speed(Fig.2),and the line-processing rate was 365 kHz(Table 1).Compared with the real-time display of the cross-sectional OCT image,the real-time en face OCTA image can guide system refraction and eye position adjustment more accurately and quickly(Fig.3,Table 2).In dynamic OCTA acquisition,the real-time display of en face OCTA images can reflect the movement of the mouse eye and its jitter,which is not evident in cross-sectional OCT images(Fig.4).In an experiment on functional retinal hyperemia,the real-time display video generated immediately after the experiment(Fig.6)can be used as a preview of the experimental results.Compared with 66.7%in the cross-sectional OCT image real-time display mode,the experimental success rate of the en face OCTA image real-time display mode was 93.3%,which proves that this mode helps avoid the situation where system adjustment and subject status problems lead to experimental failure(Table 3,Fig.5).The system can help the experimenter screen unqualified data and quickly judge the experimental results.In the future,the system could replace the frame grabber that supports a higheracquisition speedto improveits scanning speed.Conclusions We realized the real-time display of en face OCTA images in a 250 kHz SD-OCT system.Compared to the real-time display of cross-sectional OCT images using a CPU,the real-time display method in this study helps the operator adjust the system more quickly and accurately during the single acquisition of OCTA volume data and provides feedback on the subject eye state during the dynamic acquisition of OCTA volume data.The proposed real-time display method was confirmed to have a data quality feedback function in the experiment of flicker light-induced functional retinal hyperemia,which improved the experimental success rate.The line processing rate reaches 365 kHz,which can be adapted to a high-speed scanning OCTA system.