Savitzky-Golay filter based contrast-enhanced ultrasound quantification in hepatic tumors: Methodology and its correlation with tumor angiogenesis

1 Introduction

Savitzky-Golay filter was first popularized by Savitzky A and Golay M in 1964 [1], it is a digital filter that can be applied to a set of data points to smooth the data, after that improvement was made to extend the application of the filter, now it has been widely used in many fields, the process was achieved by the method of linear least squares. The predominate advantage of this filter is that while signal-to-noise ratio is increased, the signal is not greatly distorted and the width of the signal remains unchanged during the procedure [2].

The clinical practice of contrast-enhanced ultrasound (CEUS) allows for the acquisition of integrated dynamic perfusion data of organs and focal lesions, two-dimensional data sets of time and acoustic intensity of different perfusion phases are successively obtained through the recording of the examination. During the examination, respiration, circulation and digestion of the subject may lead to motion of the target organ to varying degrees, the inclination and transposition of the transducer may also move the region of interest (ROI) out of the observing plane, the out of ROI frames are the noise in the data sets. Here we introduced a Savitzky-Golay filter based CEUS quantification (SGCQ) software, by applying the theory in data smoothing, the relative maximum, minimum and the width of the data were reserved, therefore, we expect the software to reserve important perfusion characteristics while obtaining a smooth time-intensity curve (TIC). Hepatocellular carcinoma (HCC) is the most common primary malignant hepatic tumor [3]. Fifteen to twenty percent of colorectal cancer (CRC) patients present synchronous liver metastasis at the time of diagnosis, approximately 50% of CRC patients ultimately develop liver metastasis [4, 5]. This study quantitatively analyzed the perfusion of HCC and colorectal cancer metastatic liver tumor (MLT) applying SGCQ software, aiming at verifying the consistency of the software and investigating the preliminary results of the analysis.

2 Methods

The study was approved by our Institutional Ethics Committee, and written informed consent was obtained from every patient at enrollment.

2.2 Image acquisition and quantification analysis

US scans were obtained by one radiologist with over 5 years of contrast-enhanced ultrasound experience, imaging equipment was Philips iU22 system (Philips, Amsterdam, Holland) with a C5-1 convex probe (1–5 MHz). A routine US scan was performed to locate the lesion and the best plane to observe it, then a 20-gauge syringe was used for a bolus injection of 2.4 mL SonoVue (SonoVue; Bracco, Milan, Italy) through antecubital vein, followed by 5 mL saline flush. Dynamic scanning of the lesion and adjacent liver parenchyma was continuously performed of at least 150 sec to observe the dynamic enhancement, and the clip was saved as DICOM format.

Quantification analysis was performed offline on SGCQ software platform (Fig. 1a, 1b). A motion compensation calibrating region with enough gray-scale contrast was drawn on the baseline conventional screen of dual image, then motion compensation was automatically performed. The region of interest (ROI) of the lesion was defined on the frame of highest enhanced intensity in arterial phase, and another ROI was drawn on adjacent liver, avoiding artifacts, major vessel and gallbladder. Time intensity curve (TIC) was automatically fitted by the software, and the parameters and the goodness of fitting index (GOF) were presented in the form of excel, a fitting of a GOF less than 0.6 was excluded.

OPEN IN VIEWER

Fig.1

a. The interface of SGCQ software shows a HCC case, female, 47y. White arrow: motion compensation calibrating region; green arrow: ROI of the tumor; blue arrow: ROI of adjacent liver area. b. The arterial phase MRI image of the same patient. Red arrow: lesion.

2.6 Statistical analysis

OPEN IN VIEWER

Fig.2

Female, 47y, HCC. The image of CD34 antibody immunohistochemical staining section (400×magnification).

All data was presented as mean±standard deviation, t test was used to compare between two parameters in line with normal distribution and variance homogeneity, otherwise Mann-Whitney test and Satterthwaite test were used. Further diagnostic performance was evaluated for significantly different parameters. In the consistency analysis part, two-way ANOVA was used to compare the analysis of three different researchers and intraclass correlation coefficient (ICC) was calculated. Spearman correlation was used to compare the parameters with MVD count. P < 0.05 was considered to indicate a significant difference. All statistcal analysis were performed on Stata/SE 14.0 for Mac (StataCorp, Lakeway, USA).

4 Discussion

CEUS is a major development in the history of ultrasonography and is now widely used in the detection and characterization of focal liver lesions [6, 8–11]. Unlike CT and MRI, the contrast agents of ultrasound remain intravascular, and at the concentrations used in clinical scenario, the backscatter acoustic intensity of microbubbles is linearly correlates with the concentration [12, 13] and therefore reflects local vessel density. Furthermore, ultrasound contrast agent is nonnephrotoxic and with no long-term risk. The imaging process is nonradioactive, and the perfusion can be observed dynamically. The characterization of the lesions is usually made by radiologists by reviewing the exam clip, which is highly dependent on the experience they possessed in the field. Quantification software provides a more objective way to compare the hemodynamics between different lesions. SGCQ software have similar function as commercially available softwares like Sonoliver^® and Vuebox^®. For the study of focal liver lesions, respiratory motion affects the fitting of time-intensity curve, methods like respiratory gating, rejection of out-of-plane frames [14, 15] are used in some software to minimize the effect, requesting preprocessing of the data and risking possible loss of information. SGCQ software applied respiratory calibration ROI method to track the motion of ROI, allowing the analysis of integrate data, due to the algorithm of the S-G filter, the software requests relatively small program with no requirement of advanced random access memory and data processing capability of the hardware [1, 2]. The interface of the software presents both TIC curve and the parameters, including GOF (Fig. 1a), analysis could be performed multiple times of the same lesion to get a satisfying GOF. In data smoothing, the width, relative maximum and minimum of the data remain unchanged, theoretically leads to a better fitting of the curve and more precise parameters.

In the consistency analysis part, ICCs of all parameters from the tumor were greater than 0.8, indicating high retest reliability among 3 tests, ICCs of all parameters from liver areas were greater than 0.8 except for a3 and TTP. The tumor ROI was defined in arterial phase when the hypervascular pattern of the majority of the nodules [16] was obvious and the edge of the tumor was sharp, however, the ROIs of the tumor among three tests were more likely to be positioned identically among 3 researchers.

The quantification of 50 HCC tumors and adjacent liver areas showed that wash-in slope a3 and wash-out slope a2 were greater in the tumor than in the adjacent liver, suggesting that the contrast agents flowed into and out of HCC tumor faster than liver parenchyma, T0, TTP, PT of the tumor were less than that of liver parenchyma,the intensity parameters a1, PI and EI had no significant difference between the two ROIs, AUC of the tumor were less than of liver parenchyma, suggesting the total amount of contrast agents perfused into the tumor was smaller than into liver parenchyma, possibly due to the lack of portal blood supply. MTT had no significant difference between the ROIs. Diagnostic performance of the significantly different parameters a2, a3, T0 and TTP all showed an area under the receiver operating characteristic curve(AUROC) greater than 0.7, indicating intermediate diagnostic efficacy.

The quantification of 44 MLT tumors and adjacent liver areas showed that similar to HCC, a3 and a2 were greater in the tumor than in the adjacent liver, and T0, TTP, PT, AUC of the tumor were less than that of liver parenchyma and MTT had no significant difference between the ROIs. Unlike HCC, a1, PI and EI of the tumor were less than of liver parenchyma, suggesting a relatively sparse vessel density of MLT. AUROC of a2, a3, T0, TTP and AUC was greater than 0.7, indicating intermediate diagnostic efficacy.

The comparison between the two lesions showed that a3 and a2 of HCC tumor were less than of MLT tumors. AUC of HCC tumors were less than that of MLT tumors, suggesting that HCC tumors had better blood supply than MLT nudules. AUROC of AUC was greater than 0.7, indicating intermediate diagnostic efficacy. EI and PT had no significant difference between the lesions.

The processed parameters da1, dPI, and dMTT had significant difference between two lesions, dMTT of HCC was less than of MLT, da1 and dPI of HCC were greater than of MLT, AUROC of da1 and dPI were greater than 0.7.

The data simplification method was reliable in the comparison of tumors and adjacent liver areas, but in the comparison between two lesions, enhanced intensity had no significant difference while da1 and dPI were significantly different and had intermediate differential value between the lesions, suggesting that in terms of acoustic intensity parameters, the difference standardized method was more useful in the comparison between two lesions.

Among the SCGQ parameters, AUC correlated best with MVD and the correlation coefficient r = 0.72, suggesting the total amount of perfusion correlated well with the vessel density of tumor, and AUC may serve as a possible biomarker of tumor neovascularization. Wash-out slope a2 was negatively correlated with MVD (r = –0.49, P = 0.0019), the correlation was possibly due to the tortuous shape and dysfunction of tumor neovascularization, the contrast agents were detained in the microcirculation of the tumor, causing a flatter wash-out slope, the result were not proved in similar research [17], possibly relates to the dosage of contrast agent. EI was positively correlated with MVD (r = 0.66, P = 0.0000), tumors with higher MVD contains more vessel lumen in each voxel and at the time of peak intensity the voxel would contain more contrast agents, and EI may serve as a possible biomarker of tumor neovascularization, too.

Despite that a2, EI, AUC were significantly correlated with MVD, all correlation coefficients were less than 0.8, this was possibly due to the MVD count include staining units like endothelial cell or endothelial cluster, these units are not functional vessels, the perfusion and clearance of contrast agents depend on functional vessels. A study showed that MVD was not linearly correlated with the blood supply of HCC, many factors could contribute to the degree of angiographic tumor enhancement, but generally, the hypervascular type defined by angiography have greater MVD counts than the isovascular type and hypovascular type [18]. Based on the discussion above, we believed that the parameters are relevant with MVD.

Angiogenesis is a key step in tumor growth and is highly relevant with prognosis, antiangiogenetic agents like Sorafenib were often used in unresectable HCC cases [19]. MVD is a widely used marker to evaluate tumor angiogenesis and estimate prognosis [20]. In noninvasive evaluation of antiangiogenic therapy, there was no widely accept imaging biomarkers, the mechanisms of the agents are not inconsistent with response evaluation criteria of solid tumors (RECIST) [21, 22]. Zhuang et al reported that the preoperative quantification parameters, TTP, PI, AUC and wash-out time were associated with the 1-,3- and 5-years disease free survival rate [23]. Frampas et al reported that after Sorafenib treatment, a decrease in AUC of more than 40% at month 1 predicted no-progression at month 2 [24]. Studies on in vivo models also proved antiangiogenetic effect could be evaluated by quantitative CEUS [25–29]. CEUS is an ideal imaging modality for follow-ups of advanced stage HCC patients receiving both systemic therapy and locoregional therapy [30, 31], quantification parameters are obtained after simple off-line analysis, with some parameters correlate with tumor neovascularization, we expect SGCQ software to have a role in noninvasive response evaluation of antiangiogenetic therapy in HCC.

One limitation of our research is that lesions smaller than 15 mm are excluded to minimize the influence of the paratumor parenchyma perfusion to that of the tumor, another limitation is that considered the actual clinical scenario, we included tumors of depth range between 3 cm to 8 cm and inevitably included some TIC of less goodness of fitting.

In conclusion, SGCQ software shows good inter-observer consistency, wash-in slope a3 and wash-out slope a2, perfusion time are the parameters of the most diagnostic value for both HCC and MLT, AUC is useful in the analysis of MLT as well. In the comparison between HCC and MLT, AUC, da1 and dPI are of the most differential diagnostic value. Parameter a2, enhanced intensity EI and AUC correlates with tumor MVD counts, among them AUC has the maximum correlation coefficient and possibly reflects tumor angiogenesis.