Optical coherence tomography angiography (OCTA) is an important method in medical research, as it provides high-resolution imaging and detailed information about blood vessels, particularly capillaries. Since pathologies often manifest first in small vessels, OCTA is valuable for studying early lesions and holds potential in various fields of medical research, including ophthalmology^[1], dermatology^[2,3], and neurology^[4–6]. Most OCTA techniques identify the location of blood vessels by extracting information from moving objects in the sample, primarily red blood cells (RBCs). For example, phase-resolved optical coherence tomography (OCT) and Doppler OCT^[7,8] detect flow velocities by analyzing phase changes in OCT signals to visualize vessels. Correlation-based methods also leverage the motion characteristics of red blood cells (RBCs). These methods determine the location of blood vessels by comparing images of the same region taken at different times, such as the speckle variance detection method^[9] and correlation mapping method (cmOCT)^[10].

Since OCTA methods identify vessel signals by detecting motion characteristics, they have a notable limitation: there are always tail artifacts beneath the real vessel signals. When OCTA light passes through blood vessels in the sample, it is influenced by blood flow and continues to propagate deeper into the tissue. As a result, the signals from tissues beneath the vessels can exhibit motion characteristics similar to those of moving objects, leading to their misidentification as blood vessels. These tail artifacts create unclear boundaries between the base of the blood vessel and surrounding tissues, often obscuring the signals from deeper vessels. This not only degrades image quality but also complicates the reconstruction of the 3D vessel structure.

To reduce tail artifacts, many methods have been developed, most of which are image processing methods because software-based methods are more convenient. These methods include the step-down exponential filtering method^[11], transmittance effect subtraction (TAR-TES) algorithm^[12], and projection-resolved algorithm^[13]. While these methods can reduce the length of tail artifacts, they are often unable to eliminate them completely. The mean subtraction method^[14] yields better results but still struggles with clearly distinguishing between the signals and tail artifacts. Additionally, small signals in deeper areas may be missed when large vessels are located above them. Split-spectrum amplitude-decorrelation angiography (SSADA) can detect small vessels with high connectivity, but it cannot even remove the artifacts^[15]. The adaptive dynamic analysis-based method demonstrates the ability to suppress projection artifacts while enhancing micro vessel detection^[16], but this method requires long decorrelation time, resulting in extended detection time. Reflectance-based projection-resolved OCTA can remove the artifacts and has better performance in deep areas, but it is more suitable for retinal applications^[17]. Deep learning methods have gained popularity in image processing in recent years, and deep learning-based OCTA has proven to be a very effective method for suppressing blood vessel tail artifacts^[18] and providing good performance in the classification of different vessels^[19]. However, this approach requires manual labeling of vessel structures to train the network, so it is quite manpower consuming and may introduce subjective error.

In our previous study^[20], we found that structural images can significantly improve the quality of OCTA images. By identifying the locations of cavities in the structural image, we can enhance the intensity difference between vessels and artifacts, allowing for more accurate extraction of vascular signals with clearer boundaries. In this Letter, we improve the cavity signal extraction method and integrate it with other existing artifact removal techniques. Our results show that this method can effectively enhance the performance of other techniques with minimal additional steps. We demonstrate the effectiveness of our approach using a classical tail artifact removal method^[11] and a novel and effective technique^[14]. With the help of structural information, both methods can reconstruct the vascular network more clearly. The 3D imaging of palm blood vessels is presented to experimentally demonstrate our method. The results show that our method can more effectively distinguish vessels from artifacts and has better performance in deep areas.

A spectral-domain OCT (SD-OCT) system is used to acquire images. The OCT system utilizes a light source with a center wavelength of 850 nm (IPSDW0825, Inphenix), and a full width at half-maximum (FWHM) of 90 nm. The system provides an axial resolution of 4.1 µm and a transverse resolution of 6.9 µm in air. The scan range is 2.4 mm (800 pixels) × 0.6 mm (200 pixels).

Traditional OCTA methods that detect blood vessel signals yield similar results. In this study, we choose one of these methods, called the cmOCT method^[10], to acquire the original OCTA image. The cmOCT method acquires the OCTA image by calculating the correlation between OCT structure images at the same location but at different time points. It first selects a small region on the image and extracts the target information by analyzing the correlation of signals from the same region on the structure image at different time points, as shown in Eq. (1): IOCTA(x,z)=∑p=0M∑q=0N{Isy1(x+p,z+q)−E[Isy1(x,z)]}{Isy2(x+p,z+q)−E[Isy2(x,z)]}{Isy1(x+p,z+q)−E[Isy1(x,z)]}2+{Isy2(x+p,z+q)−E[Isy2(x,z)]}2,(1)where IOCTA(x,z) is an OCTA signal at an arbitrary depth position z of the xth A-line of the original image, and Isy1 and Isy2 are the structure images obtained at the same location at different time points. M and N represent the grid size and E(I) denotes the mean value of the signal magnitude within the grid. The grid is then shifted across the entire xy image to generate a 2D correlation map.

In Eq. (1), the lower the correlation between the two images, the higher the result. Regions of static tissue exhibit a high correlation, while areas containing vessels show a lower correlation, allowing us to distinguish vessels from other structures. However, in traditional OCTA images, all the vessel signals are accompanied by tail artifacts. This occurs because the signals from tissues beneath the vessels have similar characteristics with moving objects, leading to their misidentification as blood vessels.

Several methods have been developed to reduce artifacts, including the step-down exponential filtering method^[11] and the mean-subtraction method^[14].

The step-down exponential filtering method is a standard approach for removing tail artifacts in OCTA. In this method, the OCTA signal at the current point is attenuated by a factor proportional to the sum of the de-shadowed pixels above it, as follows: IOCTAsef(x,z)=IOCTA(x,z)e−[∑k=1z−1IOCTAsef(x,k)]/γ,(2)where γ is a proportionality constant that controls the rate of attenuation.

The mean-subtraction method is an effective artifact removal technique, and we also use it to demonstrate our approach. The algorithm is as follows: IOCTAms(x,z)=IOCTA(x,z)−wN×∑k=1NIOCTA(x,k),(3)where N is the total number of pixel points of the A-line, and w is a constant weighting the mean value.

Figure 1 shows the structural image and original OCTA image, along with the effects of the step-down exponential filtering method and the mean-subtraction method. In Fig. 1(a), the structural image clearly depicts blood vessels as cavity regions, whereas in Fig. 1(b), the conventional OCTA method generates pronounced tail artifacts beneath the vessels. Figures 1(c) and 1(d) demonstrate that both the step-down exponential filtering and the mean-subtraction methods effectively suppress these artifacts, with the latter showing superior performance in artifact reduction. However, even with the mean-subtraction method, the boundaries between vessels and artifacts remain unclear because the intensity of the actual vessel signals is nearly identical to that of their tail artifacts. This observation motivates our key innovation: leveraging the OCT structural image’s unique characteristic where vessels appear as cavities (unlike artifacts). To better distinguish between signals and artifacts, we first extract cavity information as follows: Icav(x,z)=C−Is(x,z),(4)where Is(x,z) describes the intensity of the OCT structure image and C is a parameter that can be adjusted based on the image intensity so that the cavity information can be obtained as clearly as possible.

To obtain a better vessel image, we need to improve the quality of the cavity image. We first compensate for depth attenuation in the B-scan structural image. Depth attenuation is influenced by the sample, and if the sample is homogeneous, the attenuation can be described by a simple scattering model. Although most samples are not perfectly homogeneous, an estimation of the global depth attenuation can already give a good result, Is′(x,z)=Is(x,z)×eαz,(5)where α is the estimation of the global depth attenuation. We choose a suitable parameter α to make the intensity of the structural image similar at different depths. We first calculate the sum of intensities at different depths in the image and select two points within the dermal layer to compute the attenuation parameter. Since the purpose of attenuation compensation is to prevent the darker regions below from being misinterpreted as cavities in subsequent processing, the parameter does not need to be highly precise; therefore, we have chosen to ignore factors such as variations in the sample surface depth, attenuation coefficient differences within the sample, and beam focus discrepancies. This approach effectively reduces the algorithm’s complexity while having minimal impact on subsequent processing steps. The B-scan structural images without and with depth attenuation compensation are shown in Figs. 2(a) and 2(b).

Although cavities appear as weak signals in the structural image, threshold segmentation cannot produce a clear cavity image due to the presence of speckle. To better extract the cavity information, we need to further enhance the difference between the cavities and the surrounding tissues. Since cavity signals are weaker than other signals, we set the signal at a point to zero when the surrounding signals are all very small, as shown in Eq. (6). This approach is more effective than setting the signal at a point to zero when the signal itself is very small, for it partially mitigates the effect of speckle, as shown in Eq. (7), $$\begin{gathered} ifIs′(x+1,z)+Is′(x,z+1)+Is′(x−1,z)+Is′(x,z−1)≤4×cL, \\ Is′(x,z)=0, \end{gathered}$$(6)ifIs′(x,z)≤cL,Is′(x,z)=0,(7)where cL is a parameter that is used to distinguish cavities and high-scattering tissues. The same operation can be used to increase the signal in high-scattering tissues: $$\begin{gathered} ifIs′(x+1,z)+Is′(x,z+1)+Is′(x−1,z)+Is′(x,z−1)≥4×cH, \\ Is′(x,z)=1, \end{gathered}$$(8)where cH is also a parameter to discover high-scattering tissues.

From Figs. 2(c) and 2(d), we can see that the difference between cavities and other high-scattering tissues is significantly larger after applying Eqs. (6) and (8). We can now use Eq. (4) to obtain a clearer cavity image, as shown in Fig. 3.

In our processing pipeline, the value of cL is the average intensity of the cavity region, and the value of cH is the average intensity of other tissue regions. C is set to approximately 1.2 times cH. The selection of these parameters is not fixed. Larger cL, smaller cH, and smaller C values tend to produce clearer images but may lead to signal loss, while the opposite reduces the effectiveness of artifact suppression. However, when these parameters vary within 1.5 times our selected values, the impact on the results is minimal. Therefore, we conclude that our method demonstrates good robustness.

After obtaining a clear cavity image, we can use it to help reduce the tail artifacts.

Since blood vessels appear as strong signals in both the OCTA and cavity images, while artifacts only appear as strong signals in the OCTA image, we can use the cavity image to suppress the artifacts in the OCTA image, as shown in Eq. (9), IOCTA′(x,z)=Icav(x,z)×IOCTA(x,z),(9)where IOCTA(x,z) describes the original OCTA image.

Now we combine the step-down exponential filtering method and the mean-subtraction method with the structural image-assisted method to further assist these methods in removing the artifacts.

A flowchart in Fig. 4 is used to outline the sequence of operations, providing a clear and intuitive representation of the processing steps in our method.

To illustrate the effect of our method, we compare the B-scan image and the en face image of the step-down exponential filtering method and the mean-subtraction method, both with and without the structural image-assisted approach, as shown in Fig. 5.

In Fig. 5, we can see that, with the help of the structural image, the distinction between vessels and tail artifacts becomes much clearer. The images show that our method can reconstruct the vessels with clear boundaries and much shorter artifacts. Using structural image measurements as the ground truth, we have added relative error analysis for vessel dimension quantification. Representative data shows that our method has 7% relative error, and the step-down exponential filtering method and the mean-subtraction method have above 100% relative error. The signal-to-noise ratio (SNR) of the image can also be improved.

To further demonstrate the effectiveness of our method, we calculate the projected images from several sequential x–z and y–z slices using the step-down exponential filtering method and the mean-subtraction method, both with and without our approach, as shown in Fig. 6.

In Figs. 6(b) and 6(c), we can see that our method can give clearer boundaries between vessels and tails. Figure 6(d) provides a unique stereoscopic demonstration of artifact suppression.

It is important to note that our method can improve the image quality, but the result depends on the effectiveness of the original method. This is because, in addition to blood vessels, other biological structures such as nerves and lymphatics can also appear as cavities in the images. Our method cannot remove the artifacts that lie both within artifact areas and the dark zones of lymphatic or neural structures. If the original method produces excessive artifacts, the residual artifacts remaining after applying our method will also increase accordingly. Nevertheless, our method introduces a new filtering criterion that enables other methods to achieve images with fewer artifacts and clearer boundaries, building upon their original results.

Of course, our method also has certain limitations. While it performs well in optimizing skin vasculature imaging, its effectiveness may be reduced in certain samples such as cerebral vasculature. The presence of large superficial vessels and the different scattering properties of brain tissue can lead to shadowing effects beneath these vessels, which may diminish the performance of our method in such scenarios.

In this Letter, we combine the structural image with the OCTA image to enhance the performance of artifact removal methods. By improving the quality of cavity images and leveraging them to differentiate between vessels and artifacts, we significantly reduce the length and prominence of tail artifacts. Vessels from the palm are used as samples to experimentally demonstrate our technique. The projected images confirm that our method effectively removes artifacts and extracts vascular signals with more accurate boundaries. The comparison of images from two typical artifact removal methods, both with and without our approach, shows that our method can reconstruct all the vessels detected by traditional OCTA artifact removal techniques, with improved performance in deeper areas.