Feasibility of a Novel Blood Noise Reduction Algorithm to Enhance Reproducibility of Ultra–High-Frequency Intravascular Ultrasound Images
Background—Ultra–high-frequency (40- to 50-MHz) intravascular ultrasound (IVUS) improves image quality compared with conventional 20- to 30-MHz IVUS. However, as the frequency of IVUS increases, high-intensity backscatter from blood components may cause visual difficulties in discrimination between the lumen and arterial wall structure. The purpose of this study was to evaluate the effect of a novel blood noise reduction algorithm (BNR) on quantitative coronary ultrasound measurements.
Methods and Results—IVUS studies using a 40-MHz transducer were performed in 35 patients with coronary artery disease. A total of 620 gray-scale images (310 pairs) were processed with and without the BNR, and lumen cross-sectional area (CSA) was determined by 2 independent observers. With the BNR, the intraobserver and interobserver correlation coefficients for lumen CSA were significantly improved (0.85 to 0.99 and 0.80 to 0.98, respectively). In the 270 images (135 pairs) in which vessel wall measurements were possible, the BNR significantly improved the intraobserver and interobserver correlation coefficients for plaque plus media CSA (0.83 to 0.99 and 0.76 to 0.97, respectively), whereas no influence was observed for external elastic membrane CSA (1.00 to 1.00 and 0.99 to 0.99, respectively).
Conclusions—This study demonstrates the feasibility of this novel algorithm to reduce blood noise, thereby enabling accurate lumen border delineation and providing reproducible measurements of both the lumen and plaque plus media CSAs. Incorporating a digital BNR may serve as an important adjunct to ultra–high-frequency IVUS imaging for improving accurate quantitative evaluation of vessel dimensions.
Intravascular ultrasound (IVUS) provides a detailed, high-quality, cross-sectional image of vessels in vivo, allowing the assessment of visual wall components and direct measurement of dimensions. IVUS is widely used as a complementary diagnostic modality to angiography for the assessment of preangioplasty and postangioplasty results,1 2 3 geometry of stent deployment, and the type and distribution of plaque contained in the vessel wall. Several studies have indicated specific IVUS parameters, such as preinterventional lesion site plaque burden and stent dimensions post deployment,4 5 to be predictive of angiographic in-stent restenosis, whereas others failed to show IVUS parameters predictive of angiographic restenosis after percutaneous coronary intervention.6 Although this is still a matter of debate,7 objective information acquired by accurate IVUS examination of target vessel segments may affect clinical practice and improve both the short- and long-term success of specific coronary intervention.8
Although improved image quality is provided by higher-frequency ultrasound transducers (40- to 50-MHz) compared with conventional (20- to 30-MHz) IVUS, a strong blood backscatter signal makes the blood appear more like scatter from soft tissue structures, potentially causing difficulties in discrimination between the lumen and arterial wall boundaries,9 10 11 especially when interpreted by less experienced personnel. Not only does the intensity of backscatter from blood increase to the fourth power of the transducer frequency, but stagnation of blood flow with rouleau formation causes a marked increase in the echo intensity of blood,12 13 both of which can significantly affect border detection in areas of disturbed flow.
To overcome these limitations, we developed and tested a novel blood noise reduction algorithm (BNR) that reduces blood scatter and is based on the fact that the speckle pattern from blood has higher spatial and temporal variations than the vessel wall.14 15 Because the BNR reduces the echo scatter from blood and allows accurate border delineation, this new technique may reduce intraobserver and interobserver variabilities in the interpretation of high-frequency IVUS imaging. The aims of this study were to validate the BNR, using IVUS images obtained in coronary arteries, by evaluating the intraobserver and interobserver variabilities of morphometric measurements and to determine the feasibility and reproducibility of this approach.
With a modified IVUS system hardware capable of BNR processing, 35 consecutive patients with symptomatic coronary artery disease were studied. IVUS was performed for diagnostic assessment of coronary atherosclerosis (n=22), after balloon angioplasty (n=7), and after stent implantation (n=6). Among patients in whom IVUS was performed for diagnostic assessment, 6 patients had severe coronary artery narrowing and required subsequent intervention. The investigation was approved by the local Committee on Human Research. All patients gave informed consent to participate in this study.
IVUS imaging was performed with a commercially available mechanical scanner and 2.6F sheath-based IVUS catheters (Discovery, Boston Scientific Corp). This IVUS catheter incorporates a single-element, 40-MHz transducer that rotates at 1800 rpm within a 15-cm-long imaging sheath. The focal zone of this transducer is between 1.5 and 2.5 mm. The overall gain setting was at the discretion of the operator and tuned to optimize vessel imaging. The transducer was withdrawn within the target segment either by mechanical pullback at 0.5 mm/s or manually to perform the imaging sequence according to the clinical situation.
The spatial/temporal filtering schematic for blood speckle suppression is illustrated in Figure 1⇓. The BNR works on the acquired envelope data before scan conversion. Digital sampling of the envelope signal for the BNR procedure was performed at a frequency of 40 MHz with 7-bit resolution. Thirty frames per second were acquired for all imaging runs.
An analytic method was developed for combining spatial and temporal filtering for blood speckle reduction in high-frequency ultrasound images. Because tissue and blood may have differing spatial frequencies, it is necessary to incorporate both temporal and spatial information rather than to rely on temporal filters alone.14 15 To combine spatial and temporal filtering, signal patterns generated by both dimensions were used. A group of 15 pixels in a target frame centered at a given pixel location was identified and likewise at the same pixel location for consecutive frames at 0.5-second intervals (15 consecutive frames) as shown in Figure 1A⇑. This group of pixels forms a 2-dimensional data matrix for a given region of interest (ROI). In the prescan conversion format, the ROI is composed of a 15×15 matrix made up of 15 pixels in the angular dimension in each of the 15 consecutive frames (Figure 1B⇑). The ROI is centered, in turn, on every point in the target frame (t0) and through every frame in the pullback sequence. The first and last 7 frames in the pullback sequence are skipped. When the edges of the center frame are encountered, the data are periodically extended by 7 pixels in the angular dimension. To determine whether the center pixel of a given ROI belongs to blood or tissue, the ROI is Fourier transformed into the frequency domain. The transformed variables now correspond to an angular spatial and temporal frequency. A ratio between the high-frequency components and low-frequency components is obtained (Figure 1C⇑). An empirically derived threshold is applied to the ratio to obtain a binary mask. In this mask, a pixel is assigned as blood speckle if the ratio is higher than the threshold; otherwise, it is assigned as tissue. The images are then modulated with the mask, with different filtering techniques applied to blood and tissue. If a pixel in the center frame (t0, S0) is labeled as blood, a minimum value of the 15 time samples (t−7 to t7, S0) is used to replace the original value. For a pixel labeled as tissue, an average with the previous output value is used. The original concept and a detailed technical description of the BNR have been reported previously.16
Validation of cross-sectional measurements by IVUS has been previously reported.1 2 3 17 The cross-sectional area (CSA) measurements included the lumen and external elastic membrane (EEM) CSAs. The luminal/intimal borders were traced manually to determine the lumen CSA. The EEM CSA, which represents the area encompassed by the medial/adventitial border, was measured by tracing the leading edge of the adventitia. Plaque+media (P+M) CSA was calculated as EEM minus lumen CSA.
Computer planimetry software (TapeMeasure, INDEC Systems) was used for morphometric image analysis. Multiple frames from direct digital signals were randomly selected for analysis. A total of 682 gray-scale images (341 pairs) were displayed with and without the BNR. Images at bifurcations (24 pairs) and aorto-ostial locations (7 pairs) precluding full delineation of the luminal boundary were excluded from analysis. For the remaining 620 images (310 pairs), the lumen CSA was measured (group 1). For the measurements of EEM CSA, 350 images (175 pairs) were excluded for the following reasons: 216 images (108 pairs) for an arc of calcium >60°,18 112 images (56 pairs) for stent shadowing, 14 images (7 pairs) because of fibrous plaque attenuation, and 8 images (4 pairs) because of severe nonuniform rotational distortion. Thus, for the remaining 270 images (135 pairs), the EEM and P+M CSAs were derived (group 2).
Each set of digitized IVUS images was analyzed by 2 independent observers. These measurements were then compared with study interobserver variability (observers 1a and 2). The images were remeasured by the first observer ≥6 weeks after the first measurement and compared to study the intraobserver variability (observers 1a and 1b).
In addition, measurements for lumen CSA on still images with BNR (A.T.) were compared with measurements reviewing the real-time imaging sequence (K.H.) to test the accuracy of the BNR.
All statistical analyses were performed with StatView statistical software, version 4.5 (Abacus Concepts). Data are expressed as mean±SD. Intraobserver and interobserver variabilities for the morphometric measurements were determined with Pearson’s correlation. The agreement of the different approaches was assessed by determining the mean±SD of the between-method differences as proposed by Bland and Altman.19 The degree of variation was presented as a coefficient of variation, defined as the SD of the between-method difference divided by the mean of the absolute value.20 A value of P<0.05 was considered to indicate statistical significance.
Feasibility of BNR
Examples of images showing mild atherosclerosis, post–balloon angioplasty disruption, and in-stent intimal hyperplasia are shown in Figure 2⇓. As can be seen, in the standard images, strong blood noise obscures visualization of the boundary between the lumen and arterial wall or neointima (Figure 2⇓, A1, B1, and C1). The BNR significantly improved border discrimination without blurring the edges or eliminating structure information contained in the vessel wall (Figure 2⇓, A2, B2, and C2). In the post–balloon angioplasty images (Figure 2⇓, B1 and B2), the BNR helped reveal a small dissection, which was difficult to appreciate by standard IVUS image presentation.
Table 1⇓ compares the results of IVUS analyses with and without the BNR. The coefficient of variation between measurements with and without BNR, expressed as the SD of the between-method differences divided by the mean of the absolute values, was consistently lower for the measurements of lumen and P+M CSA with BNR than for the measurements without BNR (Table 2⇓). The BNR had virtually no influence on the variability of measurements for EEM CSA measurements, where there was no blood interface involved.
Reproducibility of IVUS Measurements
In the 310 images for which lumen measurements were possible in both groups, the BNR significantly improved the correlation coefficient for lumen CSA from 0.85 to 0.99 for the repeated measurements by the same observer and from 0.80 to 0.98 for the measurements by 2 observers (Figure 3⇓). In addition, the BNR improved the intraobserver and interobserver correlation coefficients for P+M CSA in the 135 images in which vessel measurements were possible (Figures 4⇓ and 5⇓), whereas no influence was observed for EEM CSA. The intraobserver and interobserver correlation coefficients with the BNR (all >0.97) indicated almost perfect agreement.
To test the accuracy of the measurements with BNR compared with the real-time imaging, lumen CSA measurements of original images (310 images) were repeated by 2 observers (K.H. and A.T.). The correlation coefficient for lumen CSA between the measurements on still frames with BNR (A.T.) and measurements from real-time images (K.H.) was 0.98 (mean difference, 0.39±0.84 mm2), with a coefficient of variation of 17.4%. In addition, to examine the robust nature of this algorithm during the cardiac cycle, patients in sinus rhythm and with sequential images >5 seconds long were examined. In these 151 images (77 images in end systole and 76 images in end diastole), lumen CSA was determined 2 ways: measurements from real-time images (K.H.) and measurements made on still frames with BNR (A.T.). The correlation coefficient was 0.97 (mean difference, 0.30±0.76 mm2), and the coefficient of variation was 13.0% for end-systolic images. For end-diastolic images, both the correlation coefficient and the coefficient variation were improved (0.99 and 8.6%, respectively), and the mean difference was less (0.04±0.51 mm2 ), suggesting that the BNR may be more reliable in end diastole.
The present study has described the in vivo use of a new BNR to improve delineation of the vessel wall boundaries during IVUS imaging. The BNR reduced the high-frequency echo scatter from blood and improved delineation of internal vessel wall structures. In addition, the algorithm significantly reduced the intraobserver and interobserver variabilities of IVUS measurements of lumen area. As expected, no influence was observed on the measurements of vessel area (EEM CSA) because no blood interface was involved. To the best of our knowledge, this is the first study to determine the feasibility and reproducibility of a BNR in human coronary arteries with ultra–high-frequency imaging.
The resolution of IVUS at the current standard frequency of 30 MHz is ≈180 μm in the lateral dimension and ≈150 μm in the axial dimension. Higher-frequency transducers offer the potential for providing greater image detail yet can also produce a greatly enhanced blood signal. An in vitro study by Lockwood et al10 showed that at 50 MHz, the level of scattering measured in flowing blood was comparable to that found for fatty plaque. Using a saline flush to produce a contrast differential can be helpful to eliminate blood noise,3 but this technique is cumbersome and even potentially dangerous. Several groups have implemented filters or digital subtraction techniques to suppress blood noise.21 22 23 A recent report by Pasterkamp et al21 showed higher sensitivity and specificity for edge discrimination in mean subtraction images than in the averaged or saline flush images. However, in the mean subtraction image, information regarding the morphology of the arterial wall was lost. Li et al23 reported the feasibility of radiofrequency correlation processing to remove blood noise in the swine iliac model. In addition, a recently available online blood flow detection algorithm (ChromaFlo, Endosonics Corp)24 may also enhance the delineation of the luminal interface. The principle of this algorithm is based on time-domain processing of radiofrequency ultrasound signals. ChromaFlo is designed to detect changes in the radiofrequency signal that occur as a result of the blood flow across the ultrasound beam, highlighting the blood/intimal interface. However, flow imaging insensitivity can occur in slow-flow situation and at the vessel surface when red cell movement is minimal. Currently, ChromaFlo is limited to solid-state IVUS catheters.
In the BNR reported in this study, a 2-dimensional transform descriptor is used to characterize the difference between blood and tissue. Blood changes with each sweep of the ultrasound plane (occurring 30 times per second), whereas tissue remains reasonably stable. The result is an image that deemphasizes the blood signal (which is uncorrelated or rapidly changing in time) from the tissue signal (which is highly correlated). This BNR effectively deemphasizes the echoes from blood, even in the preinterventional images (reduced blood flow), without directly influencing the high-quality tissue pattern accomplished with high-frequency imaging. The computation time to process each frame is 7 seconds with the software implementation of the BNR in a PC-based processing system (Pentium Pro 200). It is anticipated that with firmware (digital signal processing chip), real-time imaging with the BNR is feasible, permitting direct lumen quantification in the cardiac catheterization laboratory setting.
The correlation coefficient without the BNR is low compared with previous reports.18 25 We believe this is due to enhanced blood noise by application of 40-MHz IVUS used in this study. This study demonstrated, however, that with the BNR, the intraobserver and interobserver variabilities of measurements were comparable to those of previous reports.20 25 In the present study, both intraobserver and interobserver coefficient variations were lowered with use of the BNR (<7.3% and 11.5%, respectively). A recent study by van der Lugt et al20 showed that the coefficient of interobserver variation was 17.2% and 10.5% before intervention and 11.2% and 9.2% after intervention for lumen and EEM CSA, respectively. Another study by Peters et al25 reported that the coefficient of intraobserver variation was 13.3% for lumen CSA and 8.8% for EEM CSA, whereas the coefficient of interobserver variation was 21.3% for lumen CSA and 19.4% for EEM CSA.
Reviewing the real-time imaging sequence generally aids the operator in discriminating arterial structure from blood and/or artifacts. In the present study, the measurements for lumen CSA on still frame with the BNR were compared with measurements from real-time images, showing a correlation coefficient of 0.98 and a coefficient of variation of 17.4%. This relatively large variation may be due in part to the presence of blood scattering with the real-time imaging sequence because one has to estimate the location of the contour by integrating information from neighborhood video images. Additionally, this may be a limitation of the BNR itself. The BNR requires 15 frames (0.5 seconds) to access the boundary for a given frame, which may not optimally track rapid arterial wall movement caused by catheter translation during the cardiac cycle.
Although current IVUS systems provide reasonably clear definition of intravascular pathology, interpretation can be difficult in certain patients and with specific tissues. For example, intimal hyperplasia within a stent can be difficult to image because of the very–low-intensity echoes from this tissue. High-frequency imaging shows substantial promise for improving image discrimination in these difficult cases. The main limitation of higher frequencies—the blood noise problem—appears to be resolved by the new BNR.
When morphometric measurements with ultra–high frequency IVUS are performed, the videotape has to be scrolled back and forth to elucidate the luminal border. This approach is time consuming and potentially inaccurate, especially in the cardiac catheterization laboratory, where quick and accurate measurements are crucial. In the present study, a novel BNR was implemented to facilitate visualization of the luminal boundary with 40-MHz IVUS, resulting in significantly improved reproducibility of IVUS measurements of the arterial lumen. The BNR may provide the opportunity for additional personnel, such as novice IVUS users, to measure lumen morphology quickly and reproducibly.
Recent studies have shown that the automated contour analysis system is a reliable tool for quantitative assessment of vessel dimensions in 30-MHz IVUS images.26 However, strong blood noise in ultra–high-frequency (40- to 50-MHz) IVUS images might reduce the reliability of automated contour analysis. The BNR and improved image quality provided by high-frequency IVUS will allow better distinction between vascular structure and lumen border by the use of automated 3-dimensional boundary detection and may reduce analysis time.
This study has several limitations. The first limitation for the BNR is the possible blurring effect caused by the axial and translational motion of the transducer during a cardiac cycle. However, this effect may be lessened by selecting end-diastolic images because the arterial wall motion may be minimized during this phase of the cardiac cycle. Second, the processing time for the BNR is not comparable with real-time imaging. We believe that the real-time imaging with BNR would be available in the near future, in parallel with the rapid progress in computer technology. The third limitation is the lack of a gold standard. Measurement based on images in motion were postulated as an accurate gold standard in the present study because previous studies have reported that the lumen CSA determined with IVUS correlated well with those calculated by a corresponding phantom model or pathologic sections with use of 20-MHz ultrasound transducer.17 However, a validated phantom model with flowing blood would be the only way to prove the ultimate reliability of the BNR for these morphometric measurements.
We have demonstrated the feasibility of the BNR to reduce the blood noise during routine IVUS imaging at 40 MHz. This algorithm enhances accurate border delineation and provides reproducible measurements of the lumen CSA.
Dr Hibi was supported by grants from the Getz-Stanford Cardiovascular Research Scholarship Program and the Uehara Memorial Foundation.
- Received January 25, 2000.
- Revision received May 4, 2000.
- Accepted May 11, 2000.
- Copyright © 2000 by American Heart Association
Tobis JM, Mallery J, Mahon D, et al. Intravascular ultrasound imaging of human coronary arteries in vivo: analysis of tissue characterizations with comparison to in vitro histological specimens. Circulation. 1991;83:913–926.
Nissen SE, Gurley JC, Grines CL, et al. Intravascular ultrasound assessment of lumen size and wall morphology in normal subjects and patients with coronary artery disease. Circulation. 1991;84:1087–1099.
The SH, Gussenhoven EJ, Zhong Y, et al. Effect of balloon angioplasty on femoral artery evaluated with intravascular ultrasound imaging. Circulation. 1992;86:483–493.
Peters RJ, Kok WE, Di Mario C, et al. Prediction of restenosis after coronary balloon angioplasty: results of PICTURE (Post-Intracoronary Treatment Ultrasound Result Evaluation), a prospective multicenter intracoronary ultrasound imaging study. Circulation. 1997;95:2254–2261.
Albiero R, Rau T, Schluter M, et al. Comparison of immediate and intermediate-term results of intravascular ultrasound versus angiography-guided Palmaz-Schatz stent implantation in matched lesions. Circulation. 1997;96:2997–3005.
Evans AN, Nixon MS. Biased motion-adaptive temporal filtering for speckle reduction in echocardiography. IEEE Trans Med Imaging. 1996;15:39–50.
Zhang X, Teo T-J. Method and apparatus for spatial and temporal filtering of intravascular ultrasonic image data. World intellectual property organization patent; international publication number WO 00/07035. 2000.
Crowe J, Shapo B, Stephens D, et al. Coronary artery flow imaging with an intraluminal array. In: IEEE Ultrasonics Symposium Proceedings. 1996:1481–1484.
Peters RJ, Kok WE, Rijsterborgh H, et al. Reproducibility of quantitative measurements from intracoronary ultrasound images: beat-to-beat variability and influence of the cardiac cycle. Eur Heart J. 1996;17:1593–1599.
von Birgelen C, de Vrey EA, Mintz GS, et al. ECG-gated three-dimensional intravascular ultrasound: feasibility and reproducibility of the automated analysis of coronary lumen and atherosclerotic plaque dimensions in humans. Circulation. 1997;96:2944–2952.