Noninvasive Assessment of Ki-67 Expression in Breast Cancer Using Ultrasound Radiomics: A Multi-Institutional Study

Sijie Mo; Zhibin Huang; Jing Zheng; Huaiyu Wu; Shuzhen Tang; Mengyun Wang; Jinfeng Xu; Hongtian Tian; Xiaoli Huang; Fajin Dong

doi:10.4143/crt.2025.581

Articles

Page Path: HOME > J Korean Cancer Assoc > Ahead-of print articles > Article

Original Article Noninvasive Assessment of Ki-67 Expression in Breast Cancer Using Ultrasound Radiomics: A Multi-Institutional Study: Sijie Mo¹, Zhibin Huang¹, Jing Zheng¹, Huaiyu Wu¹, Shuzhen Tang¹, Mengyun Wang¹, Jinfeng Xu¹, Hongtian Tian¹, Xiaoli Huang², Fajin Dong¹; DOI: https://doi.org/10.4143/crt.2025.581
Published online: September 5, 2025

¹Department of Ultrasound, Shenzhen People’s Hospital, The Second Clinical Medical College of Jinan University, Shenzhen, China

²Department of Ultrasound, People’s Hospital of Guangxi Zhuang Autonomous Region, Guangxi Academy of Medical Sciences, Nanning, China

Correspondence: Fajin Dong, Department of Ultrasound, Shenzhen People’s Hospital, The Second Clinical Medical College, Jinan University, Shenzhen, Guangdong 518020, China
Tel: 86-13725509696 E-mail: dongfajin@szhospital.com

^*Sijie Mo, Zhibin Huang, and Jing Zheng contributed equally to this work.

• Received: May 30, 2025 • Accepted: September 2, 2025

This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

1,813 Views
98 Download
1 Crossref

Full Article

Download PDF

Abstract

Purpose
This study aimed to develop and rigorously evaluate machine learning models using ultrasound (US) breast cancer (BC) images to predict Ki-67 expression. Additionally, the study sought to identify independent factors influencing Ki-67 expression, with further test conducted through external datasets.
Materials and Methods
This study analyzed US images of BC from 347 patients (training set: n=230; external test set: n=117) from Shenzhen People’s Hospital and Guangxi Academy of Medical Sciences. Radiomic features were extracted using manual region-of-interest delineation and the Pyradiomics package. Feature selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) and decision tree analysis, resulting in 16 features. Machine learning models—logistic regression (LR), support vector machine (SVM), and multilayer perceptron (MLP)—were developed, and their performance was assessed using the area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, and decision curve analysis. Statistical analysis included univariate and multivariate logistic regression.
Results
Three machine learning models (LR, SVM, and MLP) were developed to predict Ki-67 expression from US images. The LR model demonstrated the best diagnostic performance, with an area under the curve of 0.800 in external test set. Key predictors of Ki-67 expression included ill-defined maximum mass diameter and human epidermal growth factor receptor 2 expression, along with other significant clinical variables.
Conclusion
This dual-center study demonstrates the potential of radiomics models based on US BC images to predict Ki-67 expression accurately. As a non-invasive diagnostic tool, this approach offers valuable support for clinical decision-making and personalized treatment planning in BC patients.
Key words: Ultrasonography, Breast neoplasms, Radiomics, Machine learning, Ki-67 antigen

Introduction

According to U.S. cancer statistics, breast cancer (BC) is projected to account for approximately one-third of all new cancer diagnoses among women in 2023 [1]. Early diagnosis and treatment are crucial for improving prognosis [2]. Molecular subtyping of BC based on immunohistochemistry (IHC) enables the implementation of personalized therapeutic strategies tailored to specific tumor subtypes.

The concept of molecular classification in BC was first proposed by Dr. Perou et al. [4], and has since been refined by subsequent guidelines such as the 2013 St. Gallen International Expert Consensus, BC can be classified into four intrinsic molecular subtypes: Luminal A, Luminal B, human epidermal growth factor receptor 2 (HER2)–enriched, and triple-negative breast cancer, based on the expression of estrogen receptor (ER), progesterone receptor, HER2, and the Ki-67 proliferation index [3,4]. Notably, non-luminal subtypes tend to be more sensitive to neoadjuvant chemotherapy, while luminal subtypes—particularly Luminal B—have received relatively less attention. However, luminal tumors remain clinically significant, as Luminal B BC is associated with a worse prognosis compared to Luminal A [5]. Clinical guidelines recommend using the Ki-67 index as a key marker to distinguish between these two luminal subtypes [3].

Ki-67 is a nuclear protein expressed during active phases of the cell cycle and is closely associated with tumor cell proliferation and invasiveness. Numerous studies have identified Ki-67 as an independent prognostic factor for recurrence and survival, especially in ER-positive and HER2-negative BC [6,7]. Additionally, Ki-67 serves as a predictive biomarker for chemotherapy response, and its dynamic changes following short-term endocrine therapy may provide valuable prognostic information regarding disease-free survival (DFS) [8,9]. Therefore, as a critical marker for molecular subtyping, Ki-67 plays an essential role in guiding diagnostic classification, therapeutic decision-making, and recurrence risk assessment in patients with BC.

Core needle biopsy (CNB) or postoperative pathological examination combined with IHC remains the standard method for assessing Ki-67 expression. However, CNB is an invasive procedure that samples only a portion of the tumor tissue, which may not fully represent the entire lesion, thereby limiting the accuracy of global Ki-67 evaluation. Furthermore, whether based on CNB or surgical specimens, Ki-67 assessment by pathologists is highly dependent on individual experience and expertise, potentially leading to interobserver variability in interpretation [10]. Therefore, there is a pressing need for a more objective and reproducible approach to assist in the quantitative evaluation of Ki-67 expression. Ultrasound (US) is a more accessible modality when compared to magnetic resonance imaging (MRI) due to its real-time imaging, relatively low cost, and non-invasive nature. As a result, US has become the preferred method for BC screening in Asian women [11,12]. However, conventional imaging assessments remain largely dependent on the operator’s experience, and subtle imaging features may be missed by visual inspection alone. Radiomics, as an emerging noninvasive and quantitative approach, enables the extraction of tumor morphological, textural, and functional characteristics from medical images, facilitating the development of objective and reproducible predictive models [13-15].

Several studies have explored US-based radiomics models for predicting Ki-67 expression levels in BC [15]. Interestingly, most of these studies adopted a Ki-67 cutoff of 14% or 20% [3]. However, the International Ki-67 in Breast Cancer Working Group (IKWG) consensus recommends that only cases with Ki-67 expression < 5% or ≥ 30% should be considered for clinical decision-making regarding chemotherapy, as intermediate values (5%-30%) lack sufficient interobserver consistency and are not currently recommended for inclusion in predictive analyses [6].

Despite growing interest in US-based radiomics for predicting Ki-67 expression in BC, studies using clinically relevant thresholds (< 5% and ≥ 30%) remain scarce, especially in multicenter settings. Therefore, the primary objective of this study was to develop and evaluate machine learning models based on grayscale US images to differentiate between low (< 5%) and high (≥ 30%) Ki-67 expression in BC. These models were assessed for diagnostic performance and further validated using an external test cohort to enhance generalizability. A secondary objective was to identify independent factors associated with Ki-67 expression.

Materials and Methods

This study has been approved by the hospital ethics committee (approval No. LL-KY-2020310), and all participants were informed about the purpose and procedures of the study and provided their informed consent accordingly. The design and conduct of this research were meticulously aligned with the ethical guidelines stipulated in the Helsinki Declaration, ensuring the protection of participant rights and welfare. Additionally, this study rigorously follows the Diagnostic Accuracy Research Reporting Standards (STARD) to guarantee the comprehensiveness, transparency, and reproducibility of our diagnostic research findings [16]. These measures underscore our commitment to ethical research practices and the reliability of our reported outcomes.

1. Participants

This retrospective study continuously included all patients diagnosed with BC by postoperative pathology in two hospitals from June 2021 to June 2025, with 230 patients in the training set (age, 51.63±11.12 years) and 117 patients in the external test set (age, 49.16±8.67 years). All patients were treated according to unified criteria for selection into the group (Fig. 1). The exclusion criteria for this study were as follows: (1) absence of invasive cancer in postoperative pathology, (2) incomplete clinical data or unqualified US image quality, (3) lack of Ki-67 expression in postoperative IHC, (4) pathologically confirmed metastasis of BC to other organs, (5) presence of malignancies other than BC, (6) patients who underwent neoadjuvant chemotherapy or radiotherapy before surgery, and (7) patients with postoperative pathological IHC Ki-67 values between 5% and 30%, as well as those with HER2 overexpression. To ensure clinical relevance and alignment with IKWG recommendations, patients with HER2-positive tumors and those exhibiting intermediate Ki-67 expression (5%-30%) were excluded, as Ki-67 is considered prognostically informative for ER-positive, HER2-negative BCs with low (5%) or high (≥ 30%) proliferative indices by IKWG [6].

2. US images acquisition and segmentation

This study employed a variety of equipment, notably the GE LOGIQ E9 (GE Healthcare) and Mindray Resano 7 (Mindray Bio‑Medical Electronics Co., Ltd.), for the evaluation of US examinations. High-frequency probe was utilized in traditional US examinations to maximize resolution. The imaging parameters for patients were standardized, with gain settings adjusted to approximately 50%, and the imaging depth maintained between 3 to 6 centimeters. This approach ensured that the focus was aligned parallel to the lesion, facilitating the circumscribed visualization of each mass largest segment. All US images were stored in the Medical Digital Imaging and Communication (DICOM) format, ensuring availability for future reference and emergency use. Two radiologists (HW and HT), each with over 5 years of experience in breast US, independently delineated the regions of interest (ROIs) on stored grayscale US images using ITK-SNAP software (version 3.8.0; http://www.itksnap.org). The largest cross-section of each breast mass was selected, and the tumor margins were carefully outlined, as illustrated in Fig. 1. To evaluate inter-observer consistency in tumor segmentation, 60 US images were randomly selected and independently delineated by two experienced radiologists. The spatial agreement between the two sets of ROIs was quantified using the dice similarity coefficient (DSC), defined as:

DSC=2A∩BA+B

, where A and B represent the sets of pixels in the ROIs annotated by each radiologist. A DSC value ≥ 0.70 was considered indicative of acceptable spatial consistency. The mean DSC across the 60 cases was 0.8739±0.1511, reflecting high agreement between observers. Additionally, to assess the reproducibility of radiomic features extracted from the ROIs, the intraclass correlation coefficient (ICC) was calculated for features based on measurements from the two radiologists. An ICC greater than 0.80 was predefined as the threshold for acceptable feature reliability. Features exceeding this criterion were deemed stable.

3. Radiomic feature extraction

For the quantitative extraction of radiomic features from the delineated ROIs in US images, we utilized the Pyradiomics package ver. 3.0.1 (Python Software Foundation). This package adheres to the guidelines established by the Image Biomarker Standardization Initiative (IBSI), ensuring the standardization and reliability of the extracted features [17]. To guarantee the consistency and repeatability of the images prior to feature extraction, images were standardized using the Pyradiomics package. Furthermore, to maintain uniformity in the analysis of features such as shape and texture across different images, we resampled the voxel maximum diameter of the US images to achieve an isotropic resolution of 1 mm×1 mm. The preprocessing steps are crucial for ensuring that the analysis of radiomic features remains consistent and comparable across various images, thus enhancing the reliability of our study’s findings.

A comprehensive extraction of 1,561 radiomic features from the two-dimensional US images was conducted, spanning across seven diverse categories to offer a wide range of radiomic insights. These categories include: (1) first-order statistics, which describe the basic intensity distribution observed within the ROI; (2) Gray Level Difference Matrix (GLDM) features, which detail texture by examining the intensity differences between neighboring voxels; (3) features derived from the Gray Level Co-occurrence Matrix (GLCM), quantifying texture by analyzing the frequency of co-occurring voxel intensities; (4) Gray Level Run Length Matrix (GLRLM) features, assessing texture through the sequence lengths of voxels sharing identical intensity; (5) Gray Level Maximum diameter Zone Matrix (GLSZM) features, analyzing texture based on the maximum diameters of uniform intensity zones; (6) shape-based (2D) features that elucidate the geometric characteristics of the ROI; and (7) features from the Neighbouring Gray Tone Difference Matrix (NGTDM), describing texture by measuring intensity variation among adjacent voxels; (as illustrated in Fig. 1). These categories collectively provide a broad spectrum of insights into the radiomic landscape of the imaged area.

4. Feature selection and model building

To further ensure the stability and generalizability of the predictive models in light of the relatively limited sample maximum diameter (n=347), several strategies were employed to mitigate the risk of overfitting. In addition to feature normalization (via Z-score and mean normalization) and multicollinearity handling (by grouping features with Pearson correlation coefficients > 0.90 and retaining one representative per group), a 10-fold cross-validation strategy was adopted throughout model development. This approach helped ensure that each subset of data contributed to both training and external test, reducing model dependence on any particular partition.

In addition, the Least Absolute Shrinkage and Selection Operator (LASSO) method was employed not only for feature selection but also as a regularization technique, penalizing less informative features and preventing model complexity from escalating [18]. Decision tree analysis was subsequently applied to assess feature importance, further refining interpretability. These steps ultimately resulted in a streamlined subset of 16 US radiomics features. To enhance model interpretability and elucidate the relative importance of individual radiomic features, SHAP (SHapley Additive exPlanations) analysis was applied to the support vector machine (SVM) model. As illustrated in S1 Fig., SHAP values were used to rank the features based on their average absolute contribution to the model’s predictions.

Three machine learning models were then constructed based on the selected features: logistic regression (LR), SVM, and multilayer perceptron (MLP), as illustrated in Fig. 1. To enhance external validity, an independent validation cohort was used to assess model performance, ensuring generalizability beyond the training dataset.

5. Performance evaluation

The performance of each model was assessed comprehensively through the construction of receiver operating characteristic (ROC) curves and calculation of corresponding area under the curve (AUC), accuracy, sensitivity, and specificity. Furthermore, a detailed examination of the model’s clinical utility was conducted via decision curve analysis (DCA), which involved calculating the net benefit under various thresholds [19]. This rigorous evaluation framework ensures a robust assessment of the models’ efficacy and relevance in clinical practice.

6. Statistical analysis

Descriptive statistics were articulated through means±standard deviations for continuous variables and medians (interquartile range) along with frequencies (%) for categorical variables. The analysis of continuous variables adhered to the independent sample t-test or Mann-Whitney U test based on their distribution, while categorical variables underwent chi-squared or Fisher exact tests to discern differences. In the univariate analysis, variables that exhibited a significant association with Ki-67 expression (p < 0.05) were advanced to the multivariate LR analysis. All statistical evaluations were conducted using two-tailed tests, with a p-value threshold of less than 0.05 serving as the criterion for statistically significant differences. This approach ensures a rigorous selection of predictors for further examination in the multivariate context, enhancing the reliability and specificity of the findings in relation to Ki-67 expression.

Results

1. Clinical baseline data

The study’s methodology and patient inclusion are illustrated in Fig. 1. We incorporated 230 BC patients (average age, 51.63±11.12 years) from Shenzhen People’s Hospital along with their US images for the analysis, and an additional 117 BC patients (average age, 49.16±8.67 years) from People’s Hospital of Guangxi Zhuang Autonomous Region served as the external test group. Overall, there were no statistically significant differences in the distribution of most variables between the training and test sets, indicating good consistency (Table 1), suggesting specific heterogeneity between the two datasets in these parameters. In addition, analysis of the training set and external test set, as detailed in Table 2, revealed that distinct factors such as minimum diameter and maximum diameter (p < 0.001), posterior enhancement (p=0.010), and HER2 expression (p=0.010) were significantly associated with Ki-67 expression in both training and external test sets.

2. Radiomics feature extraction

In this research, we extracted 1,561 radiomics features from US BC images across 347 delineated ROI, covering geometric, intensity, and texture characteristics. The distribution of these particular features is shown in Fig. 1. Through dimensionality reduction via the LASSO technique, the training set was narrowed down to 33 features with non-zero coefficients. Following this, the Boruta algorithm was applied to pinpoint 16 features with potential predictive power. These selected features were subsequently used in the construction of the three models, as detailed in Fig. 1.

3. Models for differentiating Ki-67 expression using breast US

Utilizing US images of BC, we developed three distinct models to classify high and low Ki-67 expression: the LR model, the SVM model, and the MLP model. We conducted a comprehensive analysis to compare their diagnostic effectiveness. According to the findings presented in Table 3, the LR model outperformed the SVM and MLP models in diagnosing Ki-67 expression on an external test set, achieving an AUC of 0.80 (95% confidence interval [CI], 0.70 to 0.90). This was superior to the AUCs of the SVM model (0.752; 95% CI, 0.65 to 0.86) and the MLP model (0.776; 95% CI, 0.68 to 0.88). Furthermore, the LR model demonstrated the highest accuracy and specificity rates, at 0.786 and 0.893, respectively, marking it as the most effective among the evaluated models. The performance of the three models on the training set was illustrated in S2 Table.

The ROC curves for the external test sets of the three models are depicted in Fig. 1. This visualization effectively and intuitively demonstrates that the LR model exhibits superior performance in differentiating between high and low Ki-67 expression. Additionally, DCA was employed to assess the three models, with the findings also presented in Fig. 1, which indicated that the LR model provided a superior net clinical benefit across a range of threshold probabilities, highlighting its potential utility in guiding individualized clinical decision-making.

4. Correlation of clinical variables with Ki-67 expression: a univariate and multivariate analysis

In both univariable and multivariable analyses, several factors were significantly associated with Ki-67 expression. In the univariable analysis, age (p < 0.001), patient’s menstruation status (p < 0.001) as clinical variables were associated with Ki-67 expression, and among US characteristics, maximum mass diameter (p < 0.001), mass margin (p=0.010), mass posterior enhancement (p < 0.001), and mass location (p < 0.001) were also identified as factors influencing Ki-67 expression. Notably, in the multivariable analysis, only maximum mass diameter (p < 0.001), mass hypoechogenicity (p=0.038), and HER2 status (p=0.001) remained significant predictors of Ki-67 expression, indicating their critical role in differentiating high and low Ki-67 expression.

Specifically, in the multivariable analysis, mass hypoechoic characteristics were found to be negatively correlated with Ki-67 expression, with an odds ratio (OR) of 0.087 (95% CI, 0.013 to 0.603; p=0.038). Similarly, HER2 expression was also negatively associated with Ki-67 expression, with an OR of 0.298 (95% CI, 0.159 to 0.556; p=0.001). However, BC history showed a strong positive correlation with Ki-67 expression, with an OR of 4.594 (95% CI, 0.861 to 24.533), although the difference was not statistically significant (p=0.134). This suggests that while BC history appears to be associated with higher Ki-67 expression, further research is needed to clarify the significance of this relationship. These findings highlight the importance of these variables as independent predictors, as detailed in Table 4.

Discussion

BC remains a leading cause of mortality among women worldwide, underscoring the critical need for effective treatment strategies and accurate prognostic assessments [2]. The Ki-67 antigen, a key marker of tumor cell proliferation, plays a crucial role in distinguishing Luminal A and B subtypes by quantifying expression. As an important indicator of BC recurrence and prognosis, the early and accurate prediction of Ki-67 expression is essential for guiding the development of personalized neoadjuvant therapy strategies in BC patients [20].

Radiomics, a rapidly evolving field that extracts quantitative features from medical imaging data, has shown great promise in tumor characterization and biomarker assessment, including Ki-67 expression. In this study, we employed three machine learning algorithms—LR, SVM, and MLP—to develop a radiomics model based on US imaging of BC. Our results demonstrate that radiomic features derived from US imaging effectively differentiate between high and low Ki-67 expression, exhibiting strong diagnostic performance. Specifically, in the external test set, the AUC values (95% CI) for the LR, SVM, and MLP models were 0.800 (0.70-0.90), 0.752 (0.65-0.86), and 0.776 (0.68-0.88), respectively.

Notably, the LR model achieved the highest accuracy (0.786) and specificity (0.893) among the three models, in the external test set, the LR, SVM, and MLP models achieved high positive predictive values of 0.957, 0.947, and 0.939, respectively, indicating strong predictive accuracy for identifying cases with high Ki-67 expression. These findings suggest that the LR model offers superior performance in distinguishing between high and low Ki-67 expression, emphasizing its potential for clinical application. Collectively, this study highlights the utility of radiomics in enhancing BC prognostic evaluation and supports the integration of advanced machine learning approaches in clinical workflows.

The selection of Ki-67 thresholds in previous studies (e.g., 14% or 20%) has been largely empirical. For instance, the 2013 St. Gallen International Expert Consensus indicated that the choice of Ki-67 cutoffs was determined primarily by expert voting rather than rigorous evidence [3]. In contrast, the threshold values adopted in this study—Ki-67 < 5% or ≥ 30%—are based on the scientific recommendations of the International Ki-67 in IKWG [6]. Through a comprehensive analysis of multiple studies, the IKWG concluded that these thresholds represent clinically meaningful prognostic markers for BC. Although the cutoff values used in earlier studies differ from those in the present study, the findings regarding the application of radiomics to predict Ki-67 expression in BC are consistent [21,22].

A recent study also referenced the IKWG recommendations and adopted a Ki-67 cutoff of 30%. However, it did not account for the IKWG’s caution regarding the prognostic ambiguity of BC patients whose Ki-67 expression falls within the 5%-30% range. In terms of diagnostic performance, the AUC of their radiomics model based on US images was 0.688 [23]. Notably, in our study, all three radiomics models demonstrated higher AUCs in the testing cohort. This improved performance may be attributed to the inclusion of an external test set, as well as a more stable training process, which likely enhanced the generalizability of our models.

Another study reported an AUC of 0.698 for differentiating low and high Ki-67 expression using digital breast tomosynthesis (DBT) [21]. In contrast, the models developed in this study, based on US imaging, achieved superior diagnostic accuracy, with AUCs of 0.800 for LR, 0.752 for SVM, and 0.776 for MLP. The enhanced performance of the US-based models may be attributed to the inherent limitations of DBT in fully visualizing mass regions obscured by surrounding tissues, a challenge effectively mitigated by US imaging. US enables the comprehensive visualization of the entire mass without occlusion, allowing for the extraction of more detailed and accurate texture features [24]. Additionally, Kayadibi et al. [25] developed a radiomics model using MRI to predict Ki-67 expression with a cutoff value of 20%, reporting an AUC of 0.617 on the external test set. In comparison, the US-based approach in this study demonstrated significantly higher diagnostic efficacy.

Among the models developed, the LR model using US imaging exhibited the strongest capability for distinguishing between low and high Ki-67 expression. This non-invasive and effective approach provides valuable support for personalized treatment planning and clinical management of BC, underscoring the potential of US radiomics in advancing diagnostic precision and improving patient outcomes.

As a multicenter study, this research incorporated external test sets, enhancing the representativeness and generalizability of the findings while also bolstering their robustness and mitigating potential biases associated with single-center studies [26-28]. While some studies have reported a high degree of concordance between Ki-67 expression in preoperative biopsy pathology and postoperative specimen pathology, discrepancies have been observed. These differences are particularly notable in cases of highly invasive tumors, where Ki-67 expression derived from biopsies may not align with those determined postoperatively. Consequently, Ki-67 values obtained from postoperative pathology are generally regarded as more reliable. This discrepancy highlights the importance of cautious interpretation of biopsy-derived Ki-67 expression during preoperative assessments and underscores the superior diagnostic value of postoperative pathology in guiding prognostic and therapeutic strategies. A notable strength of this study is its reliance on Ki-67 expression results obtained from postoperative pathology, which provides enhanced reliability and precision compared to studies based solely on biopsy-derived data. This approach ensures a more robust foundation for evaluating Ki-67 expression and its implications for clinical decision-making.

This study investigated independent clinical predictors of Ki-67 expression in BC. Multivariable LR analysis identified maximum mass diameter and HER2 expression as statistically significant independent predictors of Ki-67 expression. Interestingly, a strong positive association was also observed between Ki-67 expression and a personal history of BC, although this relationship did not reach statistical significance in the current dataset. Given that Ki-67 is a well-established marker of cellular proliferation and tumor aggressiveness, and considering evidence that BC patients with a family history are more likely to present with higher Ki-67 indices and aggressive clinicopathological features [29], our observation of a positive trend between family history and Ki-67 expression is biologically plausible and warrants further investigation.

This study has several limitations that warrant consideration. First, the relatively small sample maximum diameters from each center may introduce bias and limit generalizability. Future studies should recruit larger, more diverse cohorts through multicenter collaborations. Another limitation of this study is its exclusive focus on intratumoral radiomic features for predicting Ki-67 expression in BC. Future work will aim to optimize imaging acquisition protocols to improve the clarity of peritumoral regions, enabling their inclusion in radiomic models to enhance predictive performance. Lastly, regarding the reliance on manual ROI segmentation in this study, although demonstrating acceptable inter-observer agreement, which may still introduce operator-dependent variability. Future studies may benefit from implementing semi-automated or AI-based segmentation approaches to improve reproducibility and operational efficiency.

This dual-center study, based on US imaging of BC, demonstrates the robust discriminative performance of three machine learning models in predicting low versus high Ki-67 expression. Among these, the LR model achieved the highest predictive accuracy, highlighting its potential clinical utility in guiding personalized treatment planning and prognostic assessment for patients with BC. Furthermore, the findings identify maximum mass diameter and HER2 expression on US as independent predictors of Ki-67 expression, offering additional diagnostic value and supporting the role of imaging-based approaches in molecular subtyping of BC.

Electronic Supplementary Material

Supplementary materials are available at Cancer Research and Treatment website (https://www.e-crt.org).

crt-2025-581_S1_Fig.pdf

crt-2025-581_S2_Table.pdf

NOTES

Ethical Statement

This study was approved by the Institutional Review Board of the Shenzhen People’s Hospital and People’s Hospital of Guangxi Zhuang Autonomous Region, specifically the Medical Ethics Committee of Shenzhen People’s Hospital and People’s Hospital of Guangxi Zhuang Autonomous Region, and all participants provided informed consent. All methods were carried out in accordance with relevant guidelines and regulations.

Author Contributions

Conceived and designed the analysis: Mo S, Huang Z, Dong F.

Collected the data: Mo S, Huang Z, Wu H, Wang M, Tian H, Huang X.

Contributed data or analysis tools: Huang Z, Zheng J, Wu H, Tang S, Wang M, Xu J, Tian H, Huang X, Dong F.

Performed the analysis: Mo S, Dong F.

Wrote the paper: Mo S.

Conflicts of Interest

Conflict of interest relevant to this article was not reported.

Funding

This project was supported by Shenzhen Science and Technology Program (GJHZ20240218114504009), Guangdong Basic and Applied Basic Research Foundation (2023A1515220148).

Fig. 1.

SHAP (SHapley Additive exPlanations) summary plot. The workflow diagram summarizes the key stages of the study, from patient inclusion and exclusion to model development and clinical utility evaluation. Region of interest (ROI) segment, tumors were manually segmented on ultrasound (US) images to delineate ROIs; Feature Selection, From ROI imaging features—including tumor morphology and texture—were extracted. To ensure feature reliability and minimize redundancy, different techniques were applied to reduce the dimensionality of the extracted features, and the Least Absolute Shrinkage and Selection Operator (LASSO) was used to select the most predictive features. Model Construction and Evaluation: The predictive performance of each model was assessed by the area under the receiver operating characteristic curve (ROC) curve. Decision curve analysis was performed to evaluate the clinical net benefit of the models. BC, breast cancer; HER2, human epidermal growth factor receptor 2; IHC, immunohistochemistry.

Table 1.

Breast cancer ultrasound findings of the study sample

Variable	Test set (n=117)	Train set (n=230)	p-value
Age (yr)	49 (42-57)	49 (45-60)	0.130^a)
Height (cm)	160 (157-162)	158 (155-162)	< 0.010^a)
Weight (kg)	60 (54-63)	58 (53-63)	0.140^a)
Menstruation status			0.090^b)
Present	55 (47.0)	132 (57.4)
Absent	62 (53.0)	98 (42.6)
BC history			0.300^b)
Absent	112 (95.7)	212 (92.2)
Present	5 (4.3)	18 (7.8)
Location			0.670^b)
Left	61 (52.1)	127 (55.2)
Right	56 (47.9)	103 (44.8)
Maximum diameter (mm)	22 (17-28)	22 (16-29)	0.930^a)
Minimum diameter (mm)	13 (11-17)	14 (10.25-18)	0.740^a)
Mass margin			< 0.010^b)
Ill-defined	84 (71.8)	195 (84.8)
Circumscribed	33 (28.2)	35 (15.2)
Orientation			0.130^b)
Parallel	19 (16.2)	55 (23.9)
Not Parallel	98 (83.8)	175 (76.1)
Hypoechogenicity			0.440^b)
Absent	14 (12.0)	20 (8.7)
Present	103 (88.0)	210 (91.3)
Posterior enhancement			0.020^b)
Absent	36 (30.8)	102 (44.3)
Present	81 (69.2)	128 (55.7)
Calcification			0.420^b)
Absent	58 (49.6)	126 (54.8)
Present	59 (50.4)	104 (45.2)
Internal CDFI			0.050^b)
Absent	15 (12.8)	51 (22.2)
Present	102 (87.2)	179 (77.8)
HER2			0.030^b)
HER2-zero	39 (33.3)	68 (29.6)
HER2-low	78 (66.7)	162 (70.4)

Values are presented as median (range) or number (%). BC, breast cancer; CDFI, color Doppler flow imaging; HER2, human epidermal growth factor receptor 2.

^a) Student’s t-test or Mann-Whitney test,

^b) Chi-square test or Fisher’s exact test.

Table 2.

Baseline characteristics of the training and test sets

Variable	Train set (low Ki-67)	Train set (high Ki-67)	p-value	Test set (low Ki-67)	Test set (high Ki-67)	p-value
Age (yr)	51.57±10.55	51.66±11.33	0.880	51.93±9.56	48.29±8.24	0.070
Height (cm)	157.84±4.19	158.03±5.23	0.780	158.96±4.97	160.54±5.23	0.320
Weight (kg)	56.48±7.28	58.94±6.90	0.130	56.54±6.24	65.85±17.95	0.020
Menstruation status			0.460			0.040
Absent	35 (62.5)	97 (55.7)		8 (28.6)	47 (52.8)
Present	21 (37.5)	77 (44.3)		20 (71.4)	42 (47.2)
Maximum diameter (mm)	21.75±9.92	24.00±11.46	0.520	21.50±16.92	24.01±10.24	0.230
Minimum diameter (mm)	12.32±6.48	15.30±5.39	< 0.001	11.64±11.93	15.97±6.61	< 0.001
Mass margin			1.000			0.100
Ill-defined	47 (83.9)	148 (85.1)		24 (85.7)	60 (67.4)
Circumscribed	9 (16.1)	26 (14.9)		4 (14.3)	29 (32.6)
Hypoechogenicity			0.460			0.060
Absent	3 (5.4)	17 (9.8)		NA	14 (15.7)
Present	53 (94.6)	157 (90.2)		28 (100)	75 (84.3)
Posterior enhancement			0.010			0.010
Absent	34 (60.7)	68 (39.1)		15 (53.6)	21 (23.6)
Present	22 (39.3)	106 (60.9)		13 (46.4)	68 (76.4)
Calcification			0.800			0.790
Absent	32 (57.1)	94 (54.0)		15 (53.6)	43 (48.3)
Present	24 (42.9)	80 (46.0)		13 (46.4)	46 (51.7)
Internal CDFI			0.980			< 0.001
Absent	13 (23.2)	38 (21.8)		11 (39.3)	4 (4.5)
Present	43 (76.8)	136 (78.2)		17 (60.7)	85 (95.5)
HER2			0.010			< 0.001
HER2-zero	8 (14.3)	60 (34.5)		10 (35.7)	29 (32.6)
HER2-low	48 (85.7)	114 (65.5)		18 (64.3)	60 (67.4)
BC history			0.610			0.160
Absent	53 (94.6)	159 (91.4)		25 (89.3)	87 (97.8)
Present	3 (5.4)	15 (8.6)		3 (10.7)	2 (2.2)
Location			0.450			0.030
Left	28 (50.0)	99 (56.9)		9 (32.1)	52 (58.4)
Right	28 (50.0)	75 (43.1)		19 (67.9)	37 (41.6)
Orientation			0.690			0.250
Parallel	15 (26.8)	40 (23.0)		7 (25.0)	12 (13.5)
Not parallel	41 (73.2)	134 (77.0)		21 (75.0)	77 (86.5)

BC, breast cancer; CDFI, Color Doppler Flow Imaging; HER2, human epidermal growth factor receptor 2; NA, not available.

Table 3.

Receiver operating characteristic curves for the test sets

Model	95% CI	Accuracy	AUC	Sensitivity	Specificity	PPV	NPV
LR	0.70-0.90	0.786	0.800	0.753	0.893	0.957	0.532
SVM	0.65-0.86	0.675	0.752	0.607	0.893	0.947	0.417
MLP	0.68-0.88	0.735	0.776	0.697	0.857	0.939	0.471

AUC, areas under the curve; CI, confidence interval; LR, logistic regression; MLP, multilayer perceptron; NPV, negative predictive value; PPV, positive predictive value; SVM, support vector machine.

Table 4.

Univariable and multivariable analyses of the train and test data set to assess variables associated with Ki-67 expression

Variable	Univariable analysis		Multivariable analysis
Variable	OR (95% CI)	p-value	OR (95% CI)	p-value
Age	1.022 (1.016-1.026)	< 0.001	0.996 (0.937-1.059)	0.912
Height (cm)	1.007 (1.006-1.009)	< 0.001	1.043 (0.992-1.096)	0.162
Weight (kg)	1.020 (1.016-1.024)	< 0.001	1.087 (1.012-1.166)	0.054
Menstruation status	3.667 (2.445-5.496)	< 0.001	0.749 (0.191-2.933)	0.727
BC history	5.000 (1.766-14.154)	0.010	4.594 (0.861-24.533)	0.134
Location	2.679 (1.861-3.857)	< 0.001	0.659 (0.285-1.528)	0.415
Maximum diameter (mm)	1.047 (1.036-1.059)	< 0.001	0.811 (0.74-0.889)	< 0.001
Minimum diameter (mm)	1.088 (1.068-1.108)	< 0.001	1.090 (0.978-1.215)	0.190
Orientation	3.268 (2.438-4.384)	< 0.001	0.610 (0.248-1.505)	0.368
Mass margin	2.889 (1.53-5.458)	0.010	0.465 (0.134-1.614)	0.312
Hypoechogenicity	2.962 (2.282-3.846)	< 0.001	0.087 (0.013-0.603)	0.038
Posterior enhancement	4.818 (3.277-7.085)	< 0.001	1.355 (0.607-3.025)	0.534
Calcification	3.333 (2.273-4.889)	< 0.001	1.971 (0.837-4.641)	0.192
Internal CDFI	3.163 (2.373-4.216)	< 0.001	0.986 (0.369-2.627)	0.980
HER2	1.716 (1.401-2.102)	< 0.001	0.298 (0.159-0.556)	0.001

BC, breast cancer; CDFI, color Doppler flow imaging; CI, confidence interval; HER2, human epidermal growth factor receptor 2; OR, odds ratio.

REFERENCES

1. Siegel RL, Miller KD, Wagle NS, Jemal A. Cancer statistics, 2023. CA Cancer J Clin. 2023;73:17–48. Article PubMed PDF
2. Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2021;71:209–49. Article PubMed PDF
3. Goldhirsch A, Winer EP, Coates AS, Gelber RD, Piccart-Gebhart M, Thurlimann B, et al. Personalizing the treatment of women with early breast cancer: highlights of the St Gallen International Expert Consensus on the Primary Therapy of Early Breast Cancer 2013. Ann Oncol. 2013;24:2206–23. PubMed PMC
4. Perou CM, Sorlie T, Eisen MB, van de Rijn M, Jeffrey SS, Rees CA, et al. Molecular portraits of human breast tumours. Nature. 2000;406:747–52. Article PubMed PDF
5. Gao JJ, Swain SM. Luminal A breast cancer and molecular assays: a review. Oncologist. 2018;23:556–65. Article PubMed PMC PDF
6. Nielsen TO, Leung SCY, Rimm DL, Dodson A, Acs B, Badve S, et al. Assessment of Ki67 in breast cancer: updated recommendations from the International Ki67 in Breast Cancer Working Group. J Natl Cancer Inst. 2021;113:808–19. PubMed PMC
7. Lee J, Lee YJ, Bae SJ, Baek SH, Kook Y, Cha YJ, et al. Ki-67, 21-gene recurrence score, endocrine resistance, and survival in patients with breast cancer. JAMA Netw Open. 2023;6:e2330961Article PubMed PMC
8. Zhang A, Wang X, Fan C, Mao X. The role of Ki67 in evaluating neoadjuvant endocrine therapy of hormone receptor-positive breast cancer. Front Endocrinol (Lausanne). 2021;12:687244.Article PubMed PMC
9. Yerushalmi R, Woods R, Ravdin PM, Hayes MM, Gelmon KA. Ki67 in breast cancer: prognostic and predictive potential. Lancet Oncol. 2010;11:174–83. Article PubMed
10. Zhu Y, Dou Y, Qin L, Wang H, Wen Z. Prediction of Ki-67 of invasive ductal breast cancer based on ultrasound radiomics nomogram. J Ultrasound Med. 2023;42:649–64. Article PubMed PDF
11. Schafgen B, Juskic M, Radicke M, Hertel M, Barr R, Pfob A, et al. Evaluation of the FUSION-X-US-II prototype to combine automated breast ultrasound and tomosynthesis. Eur Radiol. 2021;31:3712–20. Article PubMed PDF
12. Berg WA, Blume JD, Cormack JB, Mendelson EB, Lehrer D, Bohm-Velez M, et al. Combined screening with ultrasound and mammography vs mammography alone in women at elevated risk of breast cancer. JAMA. 2008;299:2151–63. Article PubMed PMC
13. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, Granton P, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer. 2012;48:441–6. Article PubMed PMC
14. Bi WL, Hosny A, Schabath MB, Giger ML, Birkbak NJ, Mehrtash A, et al. Artificial intelligence in cancer imaging: clinical challenges and applications. CA Cancer J Clin. 2019;69:127–57. Article PubMed PMC PDF
15. Lee SH, Park H, Ko ES. Radiomics in breast imaging from techniques to clinical applications: a review. Korean J Radiol. 2020;21:779–92. Article PubMed PMC PDF
16. Bossuyt PM, Reitsma JB, Bruns DE, Gatsonis CA, Glasziou PP, Irwig L, et al. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. BMJ. 2015;351:h5527.Article PubMed PMC
17. Zwanenburg A, Vallieres M, Abdalah MA, Aerts H, Andrearczyk V, Apte A, et al. The image biomarker standardization initiative: standardized quantitative radiomics for high-throughput image-based phenotyping. Radiology. 2020;295:328–38. Article PubMed PMC
18. Lund B. Logistic model selection with SAS PROC’s Logistic, Hplogistic, Hpgenselect. In: Proceedings of the 2017 Midwest SAS Users Group (MWSUG) Conference; 2017 Oct 8-10; St. Louis, MO, USA. Paper AA-02.
19. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988;44:837–45. Article PubMed
20. de Azambuja E, Cardoso F, de Castro G Jr, Colozza M, Mano MS, Durbecq V, et al. Ki-67 as prognostic marker in early breast cancer: a meta-analysis of published studies involving 12,155 patients. Br J Cancer. 2007;96:1504–13. Article PubMed PMC PDF
21. Tagliafico AS, Bignotti B, Rossi F, Matos J, Calabrese M, Valdora F, et al. Breast cancer Ki-67 expression prediction by digital breast tomosynthesis radiomics features. Eur Radiol Exp. 2019;3:36.Article PubMed PMC PDF
22. Wu Y, Ma Q, Fan L, Wu S, Wang J. An automated breast volume scanner-based intra- and peritumoral radiomics nomogram for the preoperative prediction of expression of Ki-67 in breast malignancy. Acad Radiol. 2024;31:93–103. Article PubMed
23. Liu J, Yan C, Liu C, Wang Y, Chen Q, Chen Y, et al. Predicting Ki-67 expression levels in breast cancer using radiomics-based approaches on digital breast tomosynthesis and ultrasound. Front Oncol. 2024;14:1403522.Article PubMed PMC
24. Jiang T, Song J, Wang X, Niu S, Zhao N, Dong Y, et al. Intratumoral and peritumoral analysis of mammography, tomosynthesis, and multiparametric MRI for predicting Ki-67 level in breast cancer: a radiomics-based study. Mol Imaging Biol. 2022;24:550–9. Article PubMed PDF
25. Kayadibi Y, Kocak B, Ucar N, Akan YN, Akbas P, Bektas S. Radioproteomics in breast cancer: prediction of Ki-67 expression with MRI-based radiomic models. Acad Radiol. 2022;29 Suppl 1:S116–25. Article PubMed
26. Choi SB, Park JM, Ahn JH, Go J, Kim J, Park HS, et al. Ki-67 and breast cancer prognosis: does it matter if Ki-67 level is examined using preoperative biopsy or postoperative specimen? Breast Cancer Res Treat. 2022;192:343–52. Article PubMed PMC PDF
27. Ahn S, Lee J, Cho MS, Park S, Sung SH. Evaluation of Ki-67 index in core needle biopsies and matched breast cancer surgical specimens. Arch Pathol Lab Med. 2018;142:364–8. Article PubMed PDF
28. Kim HS, Park S, Koo JS, Kim S, Kim JY, Nam S, et al. Risk factors associated with discordant Ki-67 levels between preoperative biopsy and postoperative surgical specimens in breast cancers. PLoS One. 2016;11:e0151054Article PubMed PMC
29. Mighri N, Mejri N, Boujemaa M, Berrazega Y, Rachdi H, El Benna H, et al. Association between epidemiological and clinico-pathological features of breast cancer with prognosis, family history, Ki-67 proliferation index and survival in Tunisian breast cancer patients. PLoS One. 2022;17:e0269732Article PubMed PMC

Figure & Data

REFERENCES

Citations

Citations to this article as recorded by

Utility of Ki-67 index combined with alpha-fetoprotein and lactate dehydrogenase in distinguishing mature from immature ovarian teratomas in children
Xiazhe Li, Zhiyong Zhong, Yanwei Qi, Yingxin Gong
Frontiers in Neurology.2026;[Epub] CrossRef

PubReader
ePub Link

Cite

CITE: export Copy Download Format; Close

Download Citation

Download a citation file in RIS format that can be imported by all major citation management software, including EndNote, ProCite, RefWorks, and Reference Manager.

Format:

RIS — For EndNote, ProCite, RefWorks, and most other reference management software
BibTeX — For JabRef, BibDesk, and other BibTeX-specific software

Include:

Citation for the content below
Citation and abstract for the content below

Noninvasive Assessment of Ki-67 Expression in Breast Cancer Using Ultrasound Radiomics: A Multi-Institutional Study

DOI: https://doi.org/10.4143/crt.2025.581

XML Download

Abstract
Introduction
Materials and Methods
Results
Discussion
Electronic Supplementary Material
NOTES
REFERENCES

Prognostic Significance of Bulky Nodal Disease in Anal Cancer Management: A Multi-institutional Study

Noninvasive Assessment of Ki-67 Expression in Breast Cancer Using Ultrasound Radiomics: A Multi-Institutional Study

Fig. 1. SHAP (SHapley Additive exPlanations) summary plot. The workflow diagram summarizes the key stages of the study, from patient inclusion and exclusion to model development and clinical utility evaluation. Region of interest (ROI) segment, tumors were manually segmented on ultrasound (US) images to delineate ROIs; Feature Selection, From ROI imaging features—including tumor morphology and texture—were extracted. To ensure feature reliability and minimize redundancy, different techniques were applied to reduce the dimensionality of the extracted features, and the Least Absolute Shrinkage and Selection Operator (LASSO) was used to select the most predictive features. Model Construction and Evaluation: The predictive performance of each model was assessed by the area under the receiver operating characteristic curve (ROC) curve. Decision curve analysis was performed to evaluate the clinical net benefit of the models. BC, breast cancer; HER2, human epidermal growth factor receptor 2; IHC, immunohistochemistry.

Fig. 1.

Noninvasive Assessment of Ki-67 Expression in Breast Cancer Using Ultrasound Radiomics: A Multi-Institutional Study

Variable	Test set (n=117)	Train set (n=230)	p-value
Age (yr)	49 (42-57)	49 (45-60)	0.130^a)
Height (cm)	160 (157-162)	158 (155-162)	< 0.010^a)
Weight (kg)	60 (54-63)	58 (53-63)	0.140^a)
Menstruation status			0.090^b)
Present	55 (47.0)	132 (57.4)
Absent	62 (53.0)	98 (42.6)
BC history			0.300^b)
Absent	112 (95.7)	212 (92.2)
Present	5 (4.3)	18 (7.8)
Location			0.670^b)
Left	61 (52.1)	127 (55.2)
Right	56 (47.9)	103 (44.8)
Maximum diameter (mm)	22 (17-28)	22 (16-29)	0.930^a)
Minimum diameter (mm)	13 (11-17)	14 (10.25-18)	0.740^a)
Mass margin			< 0.010^b)
Ill-defined	84 (71.8)	195 (84.8)
Circumscribed	33 (28.2)	35 (15.2)
Orientation			0.130^b)
Parallel	19 (16.2)	55 (23.9)
Not Parallel	98 (83.8)	175 (76.1)
Hypoechogenicity			0.440^b)
Absent	14 (12.0)	20 (8.7)
Present	103 (88.0)	210 (91.3)
Posterior enhancement			0.020^b)
Absent	36 (30.8)	102 (44.3)
Present	81 (69.2)	128 (55.7)
Calcification			0.420^b)
Absent	58 (49.6)	126 (54.8)
Present	59 (50.4)	104 (45.2)
Internal CDFI			0.050^b)
Absent	15 (12.8)	51 (22.2)
Present	102 (87.2)	179 (77.8)
HER2			0.030^b)
HER2-zero	39 (33.3)	68 (29.6)
HER2-low	78 (66.7)	162 (70.4)

Variable	Train set (low Ki-67)	Train set (high Ki-67)	p-value	Test set (low Ki-67)	Test set (high Ki-67)	p-value
Age (yr)	51.57±10.55	51.66±11.33	0.880	51.93±9.56	48.29±8.24	0.070
Height (cm)	157.84±4.19	158.03±5.23	0.780	158.96±4.97	160.54±5.23	0.320
Weight (kg)	56.48±7.28	58.94±6.90	0.130	56.54±6.24	65.85±17.95	0.020
Menstruation status			0.460			0.040
Absent	35 (62.5)	97 (55.7)		8 (28.6)	47 (52.8)
Present	21 (37.5)	77 (44.3)		20 (71.4)	42 (47.2)
Maximum diameter (mm)	21.75±9.92	24.00±11.46	0.520	21.50±16.92	24.01±10.24	0.230
Minimum diameter (mm)	12.32±6.48	15.30±5.39	< 0.001	11.64±11.93	15.97±6.61	< 0.001
Mass margin			1.000			0.100
Ill-defined	47 (83.9)	148 (85.1)		24 (85.7)	60 (67.4)
Circumscribed	9 (16.1)	26 (14.9)		4 (14.3)	29 (32.6)
Hypoechogenicity			0.460			0.060
Absent	3 (5.4)	17 (9.8)		NA	14 (15.7)
Present	53 (94.6)	157 (90.2)		28 (100)	75 (84.3)
Posterior enhancement			0.010			0.010
Absent	34 (60.7)	68 (39.1)		15 (53.6)	21 (23.6)
Present	22 (39.3)	106 (60.9)		13 (46.4)	68 (76.4)
Calcification			0.800			0.790
Absent	32 (57.1)	94 (54.0)		15 (53.6)	43 (48.3)
Present	24 (42.9)	80 (46.0)		13 (46.4)	46 (51.7)
Internal CDFI			0.980			< 0.001
Absent	13 (23.2)	38 (21.8)		11 (39.3)	4 (4.5)
Present	43 (76.8)	136 (78.2)		17 (60.7)	85 (95.5)
HER2			0.010			< 0.001
HER2-zero	8 (14.3)	60 (34.5)		10 (35.7)	29 (32.6)
HER2-low	48 (85.7)	114 (65.5)		18 (64.3)	60 (67.4)
BC history			0.610			0.160
Absent	53 (94.6)	159 (91.4)		25 (89.3)	87 (97.8)
Present	3 (5.4)	15 (8.6)		3 (10.7)	2 (2.2)
Location			0.450			0.030
Left	28 (50.0)	99 (56.9)		9 (32.1)	52 (58.4)
Right	28 (50.0)	75 (43.1)		19 (67.9)	37 (41.6)
Orientation			0.690			0.250
Parallel	15 (26.8)	40 (23.0)		7 (25.0)	12 (13.5)
Not parallel	41 (73.2)	134 (77.0)		21 (75.0)	77 (86.5)

Model	95% CI	Accuracy	AUC	Sensitivity	Specificity	PPV	NPV
LR	0.70-0.90	0.786	0.800	0.753	0.893	0.957	0.532
SVM	0.65-0.86	0.675	0.752	0.607	0.893	0.947	0.417
MLP	0.68-0.88	0.735	0.776	0.697	0.857	0.939	0.471

Variable	Univariable analysis		Multivariable analysis
Variable	OR (95% CI)	p-value	OR (95% CI)	p-value
Age	1.022 (1.016-1.026)	< 0.001	0.996 (0.937-1.059)	0.912
Height (cm)	1.007 (1.006-1.009)	< 0.001	1.043 (0.992-1.096)	0.162
Weight (kg)	1.020 (1.016-1.024)	< 0.001	1.087 (1.012-1.166)	0.054
Menstruation status	3.667 (2.445-5.496)	< 0.001	0.749 (0.191-2.933)	0.727
BC history	5.000 (1.766-14.154)	0.010	4.594 (0.861-24.533)	0.134
Location	2.679 (1.861-3.857)	< 0.001	0.659 (0.285-1.528)	0.415
Maximum diameter (mm)	1.047 (1.036-1.059)	< 0.001	0.811 (0.74-0.889)	< 0.001
Minimum diameter (mm)	1.088 (1.068-1.108)	< 0.001	1.090 (0.978-1.215)	0.190
Orientation	3.268 (2.438-4.384)	< 0.001	0.610 (0.248-1.505)	0.368
Mass margin	2.889 (1.53-5.458)	0.010	0.465 (0.134-1.614)	0.312
Hypoechogenicity	2.962 (2.282-3.846)	< 0.001	0.087 (0.013-0.603)	0.038
Posterior enhancement	4.818 (3.277-7.085)	< 0.001	1.355 (0.607-3.025)	0.534
Calcification	3.333 (2.273-4.889)	< 0.001	1.971 (0.837-4.641)	0.192
Internal CDFI	3.163 (2.373-4.216)	< 0.001	0.986 (0.369-2.627)	0.980
HER2	1.716 (1.401-2.102)	< 0.001	0.298 (0.159-0.556)	0.001

Table 1. Breast cancer ultrasound findings of the study sample

Values are presented as median (range) or number (%). BC, breast cancer; CDFI, color Doppler flow imaging; HER2, human epidermal growth factor receptor 2.

Student’s t-test or Mann-Whitney test,

Chi-square test or Fisher’s exact test.

Table 2. Baseline characteristics of the training and test sets

BC, breast cancer; CDFI, Color Doppler Flow Imaging; HER2, human epidermal growth factor receptor 2; NA, not available.

Table 3. Receiver operating characteristic curves for the test sets

AUC, areas under the curve; CI, confidence interval; LR, logistic regression; MLP, multilayer perceptron; NPV, negative predictive value; PPV, positive predictive value; SVM, support vector machine.

Table 4. Univariable and multivariable analyses of the train and test data set to assess variables associated with Ki-67 expression

BC, breast cancer; CDFI, color Doppler flow imaging; CI, confidence interval; HER2, human epidermal growth factor receptor 2; OR, odds ratio.