Volumetric imaging and morphometric analysis of breast tumor angiogenesis using a new contrast-free ultrasound technique: a feasibility study

This prospective study was approved by our institutional review board (IRB#: 19-003028) and was Health Insurance Portability and Accountability Act-compliant. Patients over 18 years old were referred by their primary physician to the department of Radiology for diagnostic breast imaging, including conventional ultrasound. One radiologist with 20 years of experience in breast imaging participated in this study. Moreover, the interpretation of clinical breast imaging, BI-RADS classification using grayscale ultrasound BI-RADS features as well as biopsy localization have been done by six different radiologists with 10–30 years of experience in breast imaging; however, they were not part of our investigative team and did not participate in data interpretation and or data analysis. Those patients with ultrasound-identified suspicious breast lesions recommended for biopsies were enrolled for this study. A signed written informed consent with permission for publication was obtained from each enrolled participant prior to the study. From January 2020 through December 2021, 2 males and 91 females were enrolled in this study. None of the patient population used in this study was previously reported. Participants’ information is summarized in Fig. 1. No selection bias, such as on age, sex and mass size or presence of vascularity in breast masses was present in our study eliminating pre-test probability of malignancy. All participants underwent core needle breast biopsy within an hour after the q3D-HDMI studies. Therefore, all the benign lesions were pathologically confirmed and did not need follow-up visits. A 14-gauge cutting needle was utilized. Five biopsy cores were obtained in each case. Pathology results reported by an experienced breast pathologist with more than 20 years’ experience served as the reference gold standard.

Ultrasound scanning was performed by one of two experienced sonographers with more than 30 and 15 years of clinical ultrasound scanning experience. A programmable ultrasound machine, Verasonics Vantage 256 system (Verasonics Inc., Redmond, WA), synchronized with a mechanical servo control scanning system was used for the study. A linear probe (L11-4v with central frequency at 6.25 MHz) was mounted on a motorized mechanical translation arm. The motor controller translated the linear array transducer along the elevational direction during data acquisition (Fig. 2a). Patients were asked to lie down on the back. If the lesion was located at the breast outer quadrant, a pillow was put behind the back with the patient face towards the side. It is known that peritumoral microvasculature and its morphology are important distinctive factors between malignant and benign [9, 19, 20]. Accordingly, our scanning method was designed to include the peritumoral vascularity. For this purpose, the sonographer first located the breast mass on the B-mode image and moved the probe to identify mass margins in all directions. Then, depending on the mass size, the sonographer repositioned the ultrasound probe to a few or several millimeters outside the mass margin into the normal background tissue. Starting from this point, the probe was set to automatically scan the breast in steps, covering the entire mass from one side to the other, and stopped a few to several millimeters after the mass margin has totally disappeared from the image on the opposite side. The sonographer monitored the whole scanning process to confirm that the whole mass was scanned properly. The translation step size was kept at 0.5 mm for all studies, providing a balance between acquisition time and 3D image quality. At each step, plane wave imaging was performed at a pulse repetition frequency of 550 Hz and 5 compounding angles. One thousand one hundred frames were acquired. The transducer was then placed on freeze mode for 2 s to save the IQ data. Next, the controller translated the transducer to the next scanning plane to continue data acquisition. Both the transverse and longitudinal orientations were acquired for most of the patients. For the lesions very close to the nipple, only one orientation easier for mechanical arm movement was scanned. In this study, the total scanning time was less than 5 min along each orientation. During data acquisition, the patient was instructed to remain still and breathe normally.

For each lesion, the raw data acquired at each scanning step was first processed using MATLAB (MathWorks, Natick, MA) with the q2D-HDMI technique as detailed in [6, 7]. Briefly, clutter filtering, background noise reduction and a series of morphology filtering techniques were applied to extract the microvessels. This process resulted in a set of successive 2D B-mode images and the corresponding microvessel images of lesions along the scanning direction, as shown in Fig. 2b.

The 2D images were then imported into a commercial software package, Amira 3D (Thermo Fisher, Waltham, MA), for 3D reconstruction [21]. Details for the 3D volume rendering setup are shown in Appendix 1. Figure 2b shows an example with the reconstructed microvessel image overlaid on the dilated lesion image. Finally, the 3D structure was exported as a series of 2D images for extracting quantitative parameters. To include more peripheral zone vascularity, each extracted 2D image from 3D scanning was segmented with 2 to 5 mm dilation operation, depending on size of breast mass.

MATLAB was used for quantitative analysis of the 2D images obtained from Amira 3D. The thinning algorithm as described in [7, 22] was used to extract the skeleton. Figure 2c shows the 3D skeleton image example reconstructed using MATLAB. The skeleton image was used for extracting quantitative parameters: vessel density (VD), maximum vessel diameter (D_max), mean vessel diameter (D_mean), maximum Murray’s deviation (MD_max), mean Murray’s deviation (MD_mean), maximum bifurcation angle (BA_max), mean bifurcation angle (BA_mean), microvessel fractal dimension (mvFD), number of branch points (NB), number of vessel segments (NV), mean tortuosity (τ_mean) and maximum tortuosity (τ_max). The same parameters were also obtained for q2D-HDMI by analyzing the microvessel image acquired at the scanning step, which corresponded to the B-mode image with the largest lesion cross section. The detailed list of microvessel morphological Parameters is shown in Table 1.

Statistical analysis was conducted with RStudio (R version 4.0.4, Boston, MA, USA). The Wilcoxon rank sum test was used to compare the quantitative parameters by malignancy status. A two-sided p value less than 0.05 was considered statistically significant. In cases, where the vessel has no branches, the values for BA_mean, BA_max, MD_mean and MD_max parameters are undefined, and the quantification algorithm outputs a not a number (NaN) value. Analysis with the NaN value is described in Appendix 2. All quantitative parameters, including the sub-parameters were combined using the multi-variable logistic regression analysis, and those with high p values (> 0.05) were dropped from the model as the addition of these parameters did not improve the prediction accuracy. To finally test the combined diagnostic performance of multiple morphological parameters, including the sub-parameters, leave-one-out cross-validation (LOOCV) was used. The LOOCV method has the advantage that it can compensate for the overfitting problem due to a limited sample size [23]. Malignancy probability was calculated for each lesion, and \(\mathrm{malignancy probability }={\mathrm{ logit}}^{-1}(B+\sum_{i=1}^{n}{C}_{i}{P}_{i})\), where \(B\) and \({C}_{i}\) are constants obtained from LOOCV, \({C}_{i}\) is the coefficient for the corresponding quantitative biomarker \({P}_{i}\); \(n\) is the number of quantitative biomarkers included in the prediction model; \({\mathrm{logit}}^{-1}(\alpha )\hspace{0.17em}=\hspace{0.17em}1/(1\hspace{0.17em}+\hspace{0.17em}\mathrm{exp}(-\alpha ))\). Receiver operating characteristic (ROC) curves were used to evaluate the performance of the pooled model predictions generated from the LOOCV analysis, and the optimal cut-point was defined as the point closest to the point (0,1) on the ROC curve. The corresponding 95% confidence interval (CI), sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were compared to evaluate the performance of different models. DeLong's test for paired observations [24] was also conducted based on the area under the curve (AUC) for the model performance comparison. ROC curves generated with the VD, NB, τ_mean, mvFD, MD_max, MD_mean, BA_max and BA_mean for differentiating benign from malignant lesions using q2D-HDMI and q3D-HDMI were compared. For both q2D-HDMI and q3D-HDMI, quantitative parameters VD, mvFD, τ_max, NV, NB, D_mean were included in the LOOCV model.

Of the 93 enrolled patients (mean age 52 ± 17 years, ranges from 18 to 84 years, 2 men and 91 women), 93 breast masses were examined by q3D-HDMI imaging. Table 2 summarizes the lesion information included in this study. Lesion size along the largest dimension ranged from 3 to 46 mm with a mean value of 14.7 ± 9.4 mm. Ultrasound guided core needle breast biopsy results revealed 57 benign and 36 malignant lesions. The most common benign histologic type was fibroadenoma n = 24 (42.1%). Among the malignant lesions, the most common was invasive ductal carcinoma (IDC), n = 28 (77.8%). The morphological parameters of q3D- and q2D-HDMI for benign and malignant lesions are summarized in Table 3. For both the q2D-HDMI and q3D-HDMI, the values for VD, mvFD, NV, τ_max and τ_mean were significantly higher for malignant lesions with p < 0.001. Significant differences between benign and malignant lesions were found for both the D_mean and D_max obtained from q2D-HDMI with p < 0.001 and p = 0.002, respectively. However, no significant difference was found for D_mean or D_max obtained from q3D-HDMI with p = 0.35 and p = 0.42, respectively.

The boxplots for four representative biomarkers VD, NB, τ_mean and mvFD obtained with q3D-HDMI and q2D-HDMI are shown Fig. 3. Compared to q2D-HDMI, the q3D-HDMI showed higher mean NB, τ_mean and mvFD. Also, q3D-HDMI showed lower averaged VD values for both benign and malignant lesions; this was expected as the vessel density in 2D is calculated per unit area, whereas in 3D it is calculated per unit volume. As shown in Fig. 3, ROC curves generated with a parameter obtained from q3D-HDMI demonstrated improved discrimination versus those generated with the same parameter obtained from q2D-HDMI. Significant difference was found for the ROC curve for NB with the corresponding DeLong test p values < 0.001.

The conventional B-mode and Doppler ultrasound, q2D- and q3D-HDMI images of two malignant lesions, large (as large as 36-mm) and small (as small as 3-mm) and two benign breast lesions are shown in Fig. 4 for visual comparison. The corresponding videos of the 3D rendered microvasculature images are shown in Additional files 1–4. For both the large and small malignant lesions, while 2D-HDMI shows high-definition images of tumor vascularity, the 3D-HDMI shows complete connectivity of vascularity inside and around the breast tumor. For all lesions, there are more microvessels with complete connectivity in the q3D-HDMI images. Compared to q2D-HDMI, the benign lesion vessels shown in q3D-HDMI are more regular and straight with fewer branch points, presenting less complexity when compared to those of both of large and small malignant masses, indicating the benignity of these breast lesions. The quantitative biomarkers obtained by both q3D-HDMI and q3D-HDMI are detailed in the table in row below of each breast lesion in Fig. 4 for comparison.

The ROC curves generated with MD_max, MD_mean, BA_max and BA_mean obtained from q3D-HDMI and q2D-HDMI are shown in Fig. 5a and b, respectively. All four parameters obtained from q3D-HDMI gave a significantly higher AUC when compared to the results obtained from q2D-HDMI, with p < 0.001. For both q2D-HDMI and q3D-HDMI, quantitative parameters VD, mvFD, τ_max, NV, NB, D_mean were included in the ROC Curves generated with the LOOCV model, as shown in Fig. 5c. The AUC for ROC curve of the q2D-HDMI was 83.8% (95% CI 75.2–92.3%). The corresponding sensitivity, specificity, PPV, and NPV were 86.1%, 75.4%, 68.9% and 89.6%, respectively. The AUC increased significantly (p = 0.02) for q3D-HDMI and it was 95.8% (95% CI 0.901–1.000). The corresponding sensitivity, specificity, PPV, and NPV were 91.7%, 98.2%, 97.1%, 94.9%, respectively. The negative likelihood ratio (NRL) for q2D-HDMI and q3D-HDMI were 0.184 and 0.085, respectively.

The malignancy probability [9] was calculated with LOOCV for q3D-HDMI. Figure 5d shows the malignancy probability distribution for each lesion in each BI-RADS category for q3D-HDMI. The cutoff value of the malignancy probability is 0.668, denoted by a red dashed line in Fig. 5d. As illustrated in Fig. 5d, the benign and malignant lesions in BI-RADS 3 and 4 groups were captured successfully with the q3D-HDMI. There were three false negatives (represented by red dots below the red dashed line) and one false positive (represented by blue dot above the red dashed line). Among the three false negatives, there were two ductal carcinomas in situ (DCIS), low nuclear grade cribriform type, one 9 mm, and the second was 14 mm in diameter. The third false negative case was a male patient whose breast pathology indicating a 20 mm invasive ductal carcinoma, Nottingham grade II (of III), subtype category of Luminal A (estrogen and progesterone receptors positive, HER2- negative and low levels of the protein Ki-67). The pathology of false positive was a 12 mm size fibroadenoma with lactational changes in background breast parenchyma from a female patient. The biopsy results of benign lesions also revealed 3 radial scars and 12 fibrocystic changes without hyperplasia. The prediction model of q3D-HDMI correctly predicted all three radial scars and 12 fibrocystic changes as benign. Figure 6 shows an example of a 8 mm radial scar next to a 10 mm invasive ductal carcinoma.