In optical coherence tomography angiography (OCTA), the applications of decorrelation mapping, primarily reliant on intensity data, have caught significant attention. However, this method is particularly vulnerable to the deleterious effects of noise, especially in fields characterized by low signal-to-noise ratios (SNRs). Noise artifacts have a pronounced effect on static tissue signals, which makes them exhibit elevated decorrelation between frames and in turn tends to overlap with the high decorrelation values associated with blood flow signals. This overlap detrimentally affects the quality of microvascular image acquisition. Meanwhile, classical techniques for refining decorrelation mapping, such as frequency-domain decorrelation angiography, still struggle to yield optimal results due to this inherent challenge. In response to the spurious static voxel artifacts, some studies have resorted to employing thresholding to eliminate static voxels falling below a predefined threshold. However, the global and indiscriminate nature of such thresholding often lacks a robust theoretical foundation, making the precise suppression of static voxel artifacts a complex endeavor. To this end, we present a novel OCTA approach that incorporates considerations of SNR and dynamic threshold adjustments. This innovative method is further combined with spectral analysis principles to provide a more precise means for the identification and suppression of static voxels. The ultimate objective is to enhance the microvascular imaging quality, thereby serving as a more dependable foundation for medical diagnostics.

We introduce a method for spectral amplitude decorrelation, which features dynamic threshold adjustments based on local SNRs. The methodology commences with an in-depth exploration of the complex relationship between local image SNRs and static voxels, including a comprehensive analysis of the various factors influencing this association. Subsequently, spectral analysis techniques are employed to mitigate artifacts arising from axial motion and accentuate the visualization of blood flow data. Built upon the established connection between local image SNRs and static voxels, our approach proposes adaptive thresholds for each voxel to ensure precise differentiation between dynamic and static voxels. Voxels exhibiting decorrelation values below the established threshold are categorized as static ones and subsequently suppressed. Conversely, voxels surpassing the threshold are identified as dynamic ones and are retained. Meanwhile, we further employ a sigmoid function to apply non-linear mapping to all voxels, thereby facilitating a seamless transition at the boundary between dynamic and static voxels. After the suppression of static voxels, an averaging process is applied to the decorrelation images, which allows us to reconstruct enface microvascular images by the mean projection technique. Additionally, we have established a dedicated posterior segment SS-OCT system to collect retinal data from volunteers. The effectiveness of our algorithm is rigorously validated via the data, and we conduct comparative experiments with other classical intensity-based OCTA methods to comprehensively assess its performance.

In comparison to the conventional decorrelation mapping approach, the retinal blood flow cross-sectional images processed by our algorithm exhibit prominent blood flow signals, whereas the conventional method's results are largely submerged within the noise emanating from static tissue (Fig. 6). This disparity highlights that the SSADA algorithm affected by noise-induced interference in individual spectral amplitude decorrelation images produces lower-quality enface microvascular images after averaging. In contrast, our algorithm effectively suppresses the noise arising from static voxels within individual spectral amplitude decorrelation images, ultimately yielding high-quality enface microvascular images. Compared to other intensity-based OCTA techniques, our proposed algorithm demonstrates superior performance across both high SNR skin data and low SNR retinal data, with the same preprocessing, target extraction, and image registration protocols employed. For skin data, the enface microvascular images obtained by our algorithm exhibit an SNR enhancement of approximately 4 dB in contrast to the SSADA method without static voxel suppression (Fig. 5). In the case of retinal data, our algorithm produces enface microvascular images with significantly improved contrast ratio, achieving a contrast enhancement of 0.0182 compared to the SSADA method without static suppression (Table 1).

We conduct a systematic examination of the intricate relationship between local SNRs and the decorrelation values of static voxels in OCT structural images. The results show that as noise levels on voxels increase, static voxels exhibit higher decorrelation values. Based on this pivotal finding, we introduce a dynamic threshold adjustment method within the context of spectral analysis. This combined approach adeptly leverages the sensitivity of decorrelation mapping to subtle differences and the efficacy of spectral analysis in mitigating artifacts stemming from axial motion. The retinal enface microvascular images produced by our algorithm adeptly differentiate capillaries in proximity to the macular region, underscoring the algorithm's competence in effectively suppressing static voxel noise within microvascular images. Furthermore, our algorithm consistently delivers favorable outcomes in retinal data characterized by low SNRs, resulting in enhanced image contrast ratio and superior vessel visibility. This enhancement has great potential in improving disease diagnosis and evaluation, contributing to more precise medical assessments.