In this prospective study, 228 single breast lesions from 228 patients were identified via the routine ultrasonic examination and classified as category 4 according to ACR US-BI-RADS (Figure 1). All patients were advised to undergo a CEUS examination before US-guided coarse needle biopsy (CNB) and surgery. However, 43 patients did not perform the CEUS examination and underwent surgery directly. A total of 33 patients failed to follow up and did not obtain their pathological results. Finally, 152 breast lesions from 152 patients got both histopathological results and CEUS videos successfully, and all the patients gave informed consent. The research institute granted ethical approval (No. 2018200) to carry out the study within its facilities. In the process of data analysis, two patients were excluded for failing to acquire high-quality CEUS data. Finally, 150 breast lesions from 150 patients were analyzed in this study (Figure 1). Of the 150 suspicious breast lesions, 101 (67.3%) were malignant (16 BI-RADS 4a, 31 BI-RADS 4b, 54 BI-RADS 4c) and 49 (32.7%) were benign (35 BI-RADS 4a, 12 BI-RADS 4b, 2 BI-RADS 4c) (Table 1).

Using the Siemens Acuson Sequoia 512 color Doppler ultrasound diagnostic system, the routine US examination was performed with an 18L6 linear transducer of frequency of 4.6–17.8 MHz, and CEUS was performed with a 10L4 linear transducer of frequency of 2.9–9.9 MHz. Patients were instructed to keep a supine position and expose the breast sufficiently. Ultrasonographers switched the routine model to the contrast model after finding a lesion. The contrast agent, 4.8 mL of SonoVue (Bracco Inc., Milan, Italy), was injected into the peripheral vein of the patient. The dynamic contrast-enhanced process within the lesion was observed and recorded for 2 min. The video data were stored automatically on the machine’s hard disk. Based on the grayscale images and the contrast-enhanced videos, we recorded the location, size, shape (regular/irregular), boundary (clear/unclear), blood types, echogenic foci (macro/micro-calcifications), and axillary lymph nodes of lesions. We also recorded the perfusion patterns and the direction of contrast entering the nodules, the enlarged size and the shape of enhanced lesions, and the enhanced vessels within and surrounding lesions. Finally, the DICOM format of the contrast videos was input into a quantitative software called TomTec SonoLiver v1.1.15.0, where the region of interest (ROI) of normal breast gland tissue and lesions (including the whole lesions and partial lesions with solid components) was marked and analyzed. The software provided five parameters based on its default smoothing curves, including: quality of fit (QOF, which measures the perfusion curve fitting degree), maximum intensity (IMAX, the echo intensity of the contrast medium at its peak), rising time (RT, the time from the onset of the contrast medium and its peak), time to peak (TTP, the time between the start of the contrast agent injection and its peak), and mean transit time (mTT, the mean local transition time of contrast media). The results of the software output in this study were completed by two highly trained radiologists. When the two radiologists disagreed, a third radiologist with over 10 years of experience made the final decision.

The qualitative features were compared via the chi-square and Fisher’s exact test. The quantitative parameters were compared via a t-test, Bartlett’s test, ANOVA test, and Kruskal test. A p < 0.05 was considered statistically significant. Single-factor and multi-factor logistic regression with forward stepwise analysis were applied to screen for independent risk factors and establish an integrated model to identify benign and malignant lesions as well as histopathological subtypes of suspicious breast lesions. The ROC curve was used to evaluate the integrated model and calculate the area under the curve (AUC). The cutoffs for the qualitative features and quantitative parameters of CEUS were determined, and the corresponding sensitivity and specificity were calculated based on the maximum value of Youden’s index. The DeLong test was used to compare the differences between the AUCs of the ROC curves. All statistical analyses were executed using R software version 4.0 (R codes are shown in the Supplementary material).

Of 150 breast lesions scored as category 4 of US-BI-RADS, 101 were breast cancers (16 BI-RADS 4a, 31 BI-RADS 4b, and 54 BI-RADS 4c) and 49 were atypical benign lesions (35 BI-RADS 4a, 12 BI-RADS 4b, and 2 BI-RADS 4c). The average age of the patients with malignancies and benignities was 52.8 and 44.6, with standard deviations of 13.5 and 10.9, respectively. The 49 benign lesions included 20 fibroadenomas (FIB) with atypical grayscale ultrasonic features, 10 hyperplastic nodules, 1 hyperplastic nodule with FIB, 4 intraductal papillomas, and 14 mass-like non-puerperal mastitis (NPM). The 101 malignant lesions included 83 invasive ductal carcinomas (IDC), 12 ductal carcinomas in site (DCIS), 2 invasive lobular carcinomas, 2 intraductal papillary carcinomas, and one mucinous carcinoma (Table 1). The average size of 49 benign lesions is 18.2 mm with a standard deviation of 12.1 mm, and the average size of 101 malignant lesions is 23.2 mm with a standard deviation of 13.3 mm (Supplementary Table 1). Other data, including age, menopausal history, family history, location of lesions, lymphatic metastasis, and ultrasound and contrast-enhanced ultrasound characteristics, are shown in Supplementary Table 1.

Among the risk qualitative features of CEUS, enlarged scope, heterogeneous perfusion, perfusion deficiency, crab clam-like enhancement, more than two enhanced vessels within lesions, and surrounding enriched vessels inserting into lesions represented statistically significant differences for differentiating benign from malignant lesions (p < 0.01). Of these features, crab clam-like enhancement had the highest specificity of 95.9%, but the lowest sensitivity of 51.5%, with moderate accuracy of 66.0% (Table 2). The surrounding enriched vessels inserting into lesions had the highest sensitivity of 100% and the highest accuracy of 87.3%, though with moderate specificity of 61.2% (Table 2).

In a comparison of CEUS qualitative features for differentiating the four histopathological subtypes (IDC, DCIS, FIB, and NPM), except the directional perfusion, other risk features showed significant overall differences (p < 0.01) (Supplementary Table 2). In the paired comparison, the crab clam-like enhancement, more than two enhanced vessels within lesions, and surrounding enriched vessels inserting into lesions could differentiate IDC and DCIS from atypical FIB and NPM (p < 0.05) (Figures 2, 3) but showed no significant difference between IDC and DCIS, or between FIB and NPM (p > 0.05) (Figure 2). The enlarged scope, irregular shape, and perfusion deficiency were valuable in differentiating FIBs from the others (p < 0.05), although they showed no significant difference among the paired comparison of IDC, DCIS, and mass-like NPM (p > 0.05) (Figure 2). The heterogeneous perfusion showed a significant difference only between IDC and atypical FIB (p < 0.05) (Figure 2).

After the quantitative analysis, 140 of 150 breast lesions had QOFs over 50% and the QOFs values were higher in the whole enhanced lesions than in the normal enhanced breast gland tissue and partially enhanced lesions with solid composition (Supplementary Figure 1). In a comparison of four quantitative parameters (IMAX, RT, TTP, and mTT) of CEUS in differentiating the benign from malignant breast lesions, only IMAX was significantly higher for the malignancy than the benignity in both whole and part of enhanced lesions (p < 0.05). The other three parameters showed no significant difference between malignancies and benignities (Supplementary Figure 2 and Supplementary Table 2). IMAX showed the highest diagnostic value for differentiating the benign from malignant breast lesions for the whole lesions (AUCROC: 0.62, 95% confidence interval: (0.52: 0.72), the cutoff value of 299.4%, sensitivity: 44.8%, specificity: 77.3%, negative predictive value: 39.1%, positive predictive value: 81.1%, accuracy: 55%, p < 0.05), and the partial lesions with solid components (AUCROC: 0.61, 95% confidence interval: (0.51: 0.71), cutoff value: 695.6%, sensitivity: 60.4%, specificity: 65.9%, negative predict value: 43.3%, positive predict value: 79.5%, accuracy: 62.1%, p < 0.05) (Table 3 and Supplementary Figure 2). In the paired comparison of the histopathological subtypes of breast lesions, none of the quantitative metrics showed a significant difference in differentiating between IDCs, DCISs, FIBs, and NPMs (p > 0.05) (Supplementary Figure 3 and Supplementary Table 4).

In the establishment of the logistic regression analysis model, heterogeneous perfusion, crab clam-like enhancement, and partial_IMAX were independent risk factors for benign and malignant breast lesions (p < 0.05). Of these features, the crab clam-like enhancement had the highest OR of 30.91. Heterogeneous perfusion and partial_ IMAX OR values are 5.46 and 1.01, respectively (Table 4).

The diagnostic efficacy of the integrated model for suspicious breast lesions is higher than that of independent risk factors (heterogeneous perfusion, crab clam-like enhancement, and partial_ IMAX), with an AUCROC of 0.89 (95% CI: 0.83–0.94). When the cutoff value was 0.608, the sensitivity and specificity were 83.0 and 78.0%, respectively, and the accuracy was 81.3% (Supplementary Table 5 and Figure 4).

In diagnosing benign pathological subtypes of breast lesions, there was no significant difference between independent risk factors and integrated model for diagnosis of FIBs and NPMs (p > 0.05). In the diagnosis of malignant pathological subtypes of breast lesions, there was no significant difference between independent risk factors and integrated model in the diagnosis of IDCs and DCISs (p > 0.05) (Supplementary Table 5).