The retrospective study was approved by the institutional review board at our hospital. The inclusion criteria were: (1) Patients who underwent both ABVS and SE examinations at our hospital between August 2019 and May 2022 and subsequently underwent biopsy or surgical resection within two weeks with a pathologically confirmed diagnosis. (2) Patients with complete imaging data. (3) Patients’ breast lesions were hypoechoic solid lesions. The exclusion criteria were: (1) Patients whose images were of poor quality; (2) Patients who underwent aspiration or clinical treatment before examining target lesions. Eventually, we included 423 solid hypoechoic breast lesions from 423 female patients. By generating random numbers, they were allocated into training and validation groups in the ratio of 7:3. The flow is shown in Figure 1 .

In this study, all images used were obtained using the ACUSON S2000 US machine and its accompanying ABVS system.

The ABVS and SE images were sequentially imported into 3D Slicer 5.2.1 for image processing, manual segmentation of the ROI and extraction of radiomics features. We delineated ROI for lesions in both SE and ABVS images. Notably, in the ABVS images, we delineated the ROI on the coronal, sagittal, and transverse planes of the lesions, respectively. Further details can be found in Figure 2 . This procedure was performed with the participation of two physicians. Physician A, with five years of experience in ultrasound-based breast disease diagnosis, performed outlining for all the lesions. Physician B, with eight years of experience in ultrasound-based breast disease diagnosis, conducted lesion outlining on the training group to validate ROI outlining reproducibility. Then, we utilized the Pyradiomics package within 3D Slicer to extract radiomics features from the SE and ABVS images, respectively. The features extracted encompassed first-order statistics features, texture features (including the gray level co-occurrence matrix (glcm), gray level dependence matrix (gldm), gray level run length matrix (glrlm), gray level size zone matrix (glszm), and the neighbouring gray tone difference matrix (ngtdm)), as well as post-wavelet transformed features. Subsequently, we subjected the extracted features to screening. By utilizing the intra-class correlation coefficient (ICC) analysis, we can identify features that exhibit high levels of reproducibility (ICC > 0.75) (32), subsequently, radiomics features extracted from the region of interest segmented by physician A were utilized for further analysis. All feature values were normalized using Zscore. The radiomics features of both modalities were subjected to dimensionality reduction through the Mann-Whitney U test and least absolute shrinkage and selection operator (LASSO) regression and to identify features with strong qualitative diagnostic ability for solid hypoechoic breast lesions.

This study performed a univariate analysis in order to determine the risk predictive variables associated with breast cancer(P<0.05), which were then combined with the results of conventional imaging assessment and radiomics analysis. Based on these findings, we integrated relevant clinical risk factors, the Radscore, and the BI-RADS category of lesions to develop a nomogram for assessing the malignancy risk in such breast lesions by multivariate logistic regression analysis. Subsequently, the nomogram’s diagnostic performance was compared to that of the BI-RADS model and SE+ABVS radiomics model. To evaluate the diagnostic performance of the models, we calculated the area under the receiver operating characteristic curve (AUC) for each model in the training group, validation group, and in the BI-RADS category 4 lesions within both groups. Furthermore, the DeLong test was used to examine differences in AUC values between different models. The nomogram’s goodness of fit was investigated graphically and by calculating significance by plotting the calibration curve and conducting the Hosmer-Lemeshow test. Lastly, clinical decision analysis curves were drawn for quantifying the net benefits of the BI-RADS model, SE+ABVS radiomics models, and nomogram at various threshold probabilities.

SPSS 23.0, R 4.2.2, and MedCalc 19.6.0 were utilized for statistical analysis and graph plotting. The ‘psych’, ‘survival’, ‘glmnet’, ‘rms’, ‘ResourceSelection’, and ‘rmda’ packages were used in R. We performed normality tests on each group of data and selected the appropriate hypothesis test based on the results to compare the distribution of data between the training and validation groups. The study has chosen a significance level of 0.05 as the threshold for detecting statistical differences.

The study included 423 breast lesions that were pathologically confirmed to include 215 benign lesions and 208 malignant lesions. Table 1 demonstrates that both the clinical basis data and sonographic features of lesions were evenly distributed in the training and validation groups, indicating no statistically significant differences between the two groups (P>0.05). Furthermore, the univariate risk analysis revealed that patients with malignant lesions had significantly higher age and lesion’s maximum diameter compared to those with benign lesions in both groups (P<0.05). However, no correlation was seen between the location of the lesion and the malignant risk of the lesion (P>0.05). Hence, we regarded age and lesion size as predictor variables in the context of breast cancer. Regarding the sonographic features of the lesions, there were statistically significant differences (P<0.01) observed in morphology, borders, margins, orientation, microcalcifications, retraction condition of the coronal plane, and elasticity scores between benign and malignant lesions within both groups. while no statistical differences were found in posterior echogenicity (P>0.05). This study assessed the malignancy risk of lesions by these sonographic features of them, and the BI-RADS categories obtained were also significantly different in benign and malignant lesions (P<0.01).

The SE and ABVS images generated 837 and 2511 radiomics features (ABVS cross plane, sagittal plane, and coronal plane, each generated 837 features), respectively. The training group’s SE radiomics features as well as the ABVS coronal plane, transverse plane, and sagittal plane radiomics features underwent sequential ICC analysis, Mann-Whitney U test, LASSO regression analysis with tenfold cross-validation for dimensionality reduction. Finally, a total of 14 features were selected, comprising four SE radiomics features and ten ABVS radiomics features (two from the coronal plane, three from the transverse plane, and five from the sagittal plane). All of these radiomics features are texture features, one of which was from the original image and thirteen were obtained after wavelet transform ( Figure 3 ).

By comparing and validating the diagnostic efficacy of the radiomics models ( Figure 4 , Table 2 ), the AUC values of the selected ABVS and SE features for distinguishing between benign and malignant solid hypoechoic breast lesions were consistently above 0.8 in both the training and validation groups. Moreover, compared to any single-modality radiomics models, the SE+ABVS radiomics model, which integrated the radiomics features of two imaging modalities, demonstrated significantly higher AUC values in both training (All P<0.01) and validation groups (compared to the ABVS radiomics model: P<0.01, compared to SE radiomics model: P<0.05). These outcomes suggest that combining radiomics features from both SE and ABVS could enhance the accuracy of diagnostic models. Thus, the Radscore for each patient was obtained by weighting the corresponding coefficients for each feature in the SE+ABVS radiomics model., the formula is shown below, the Radscore for malignant lesions was found to be significantly higher than that for benign lesions within both groups. (Training group: 2.86 + 2.66, -2.34 + 1.80, P<0.01; Validation group: 2.60 + 2.31, -2.17 + 1.85, P<0.01).

Based on the clinical risk factors identified through univariate analysis, BI-RADS categories determined from imaging assessments, and Radscore obtained from radiomics analysis, we constructed a nomogram using multivariate logistic regression to visually assess the risk of malignancy in solid hypoechoic breast lesions, the nomogram incorporated the patient’s age, lesion’s maximum diameter, Radscore, and BI-RADS category. As illustrated in Figure 5 , Radscore had the highest weightage followed by BI-RADS score while age and maximum diameter of the lesion exerted less influence on assessment results.

Figure 6 and Table 3 present that the BI-RADS model, SE+ABVS radiomics model, and nomogram are effective in predicting the malignancy risk in solid hypoechoic breast lesions, Notably, the nomogram exhibits superior diagnostic performance with higher AUC values (0.972, 0.964) in training and validation group compared to both the BI-RADS model (AUC: 0.930, 0.916) and SE+ABVS radiomics models (AUC: 0.941, 0.933). Furthermore, its difference with BI-RADS model and SE+ABVS radiomics model was statistically significant in both groups (P<0.05). Besides, we further compared the diagnostic efficacy of the three models for BI-RADS category 4 lesions within the two groups. The results revealed that the nomogram (AUC: 0.952, 0.930) consistently exhibited higher AUC values than both the BI-RADS model (AUC:0.844, 0.839) and SE+ABVS radiomics model (AUC:0.915, 0.899). Moreover, there were consistently statistically significant differences between the nomogram and BI-RADS model (All P<0.01). However, in comparison to the SE+ABVS model, the nomogram was only statistically different from it in the training group (P<0.05), but not in the validation group (P>0.05). Other than that, in terms of diagnostic sensitivity, specificity, and accuracy, Although the specificity of the nomogram was slightly inferior to that of the BI-RADS model in the training group, it significantly improved diagnostic sensitivity. Furthermore, its diagnostic parameters were at the highest level across all validation groups. These results suggest that the nomogram exhibited the best overall diagnostic performance. Finally, we observed that the AUC values of the SE+ABVS Radiomics model consistently outperformed those of the BI-RADS model, and a statistically significant difference was found between them when diagnosing BI-RADS category 4 lesions of the training group (P<0.05). This finding highlights the ability of radiomics analysis to detect deep-seated features within the images, ultimately leading to improved diagnostic efficiency.

The calibration curve exhibits a favorable fit of the nomogram ( Figures 7A, B ). indicating that the predicted risk by the nomogram was close to the observed risks. The results from the Hosmer-Lemeshow test further proved that the differences between them did not present statistical significance in either the training group (P=0.70) or validation group (P=0.95).

The clinical decision analysis curve ( Figure 7C ) indicates that utilizing the BI-RADS model, SE+ABVS radiomics model, and nomogram for decision-making significantly improved the net benefit for patients compared to the assumption of intervention for all lesions or no intervention at all. Furthermore, the nomogram provided a greater net benefit to patients compared to both the BI-RADS model and SE+ABVS radiomics model.