Ultrasonographic morphological characteristics determined using a deep learning-based computer-aided diagnostic system of breast cancer.

ABSTRACT: To investigate the correlations between ultrasonographic morphological characteristics quantitatively assessed using a deep learning-based computer-aided diagnostic system (DL-CAD) and histopathologic features of breast cancer.This retrospective study included 282 women with invasive breast cancer (<5 cm; mean age, 54.4 [range, 29-85] years) who underwent surgery between February 2016 and April 2017. The morphological characteristics of breast cancer on B-mode ultrasonography were analyzed using DL-CAD, and quantitative scores (0-1) were obtained. Associations between quantitative scores and tumor histologic type, grade, size, subtype, and lymph node status were compared.Two-hundred and thirty-six (83.7%) tumors were invasive ductal carcinoma, 18 (6.4%) invasive lobular carcinoma, and 28 (9.9%) micropapillary, apocrine, and mucinous. The mean size was 1.8 ± 1.0 (standard deviation) cm, and 108 (38.3%) cases were node positive. Irregular shape score was associated with tumor size (P < .001), lymph nodes status (P = .001), and estrogen receptor status (P = .016). Not-circumscribed margin (P < .001) and hypoechogenicity (P = .003) scores correlated with tumor size, and non-parallel orientation score correlated with histologic grade (P = .024). Luminal A tumors exhibited more irregular features (P = .048) with no parallel orientation (P = .002), whereas triple-negative breast cancer showed a rounder/more oval and parallel orientation.Quantitative morphological characteristics of breast cancers determined using DL-CAD correlated with histopathologic features and could provide useful information about breast cancer phenotypes.

Introduction

Breast cancer is the most commonly diagnosed cancer and the leading cause of cancer-related deaths in women worldwide.[ Histopathologic evaluation of breast cancer by tissue sampling is essential for treatment planning and prediction of prognosis, and it provides information on the tumor size, histologic grade, nodal status, expression of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2), which are very important prognostic markers.[ Recent advances in image acquisition, computational power, and algorithmic development have allowed quantitative information acquisition from computer-aided diagnosis.[ Imaging techniques have been shown to non-invasively provide information regarding the underlying histopathology,[ and this technology has been adopted previously to phenotypically characterize breast cancers using preoperative magnetic resonance imaging (MRI).[

Ultrasound (US) has many advantages such as ease of accessibility without the need for radiation and contrast material. Recently, high-resolution US has proven useful for the evaluation of small structures, such as nerves or tendons, and its diagnostic accuracy is comparable to that of MRI.[ US is a widely used imaging modality for breast cancer detection in adjunct to mammography (MG) in women with dense breasts and breast mass differentiation.[ In breast cancer evaluation, US can be used to assess the extent, multifocality, and multicentricity of the tumor in the breast, and axillary lymphadenopathy.[ Besides, US is useful for the treatment response monitoring of breast cancer.[ However, US has some limitations, including operator dependency and limited reproducibility. Consequently, quantitative and objective analyses of morphological characteristics with US are limited, and the correlation between breast cancer and histopathologic features has not been well investigated.

Recent advances in artificial intelligence, particularly deep learning algorithms, have gained extensive attention owing to their excellent performance in image recognition tasks.[ It can be used to detect subtle findings in US images that expert radiologists overlook and automatically produce a quantitative assessment.[ Deep learning-based computer-aided diagnosis (DL-CAD) software for breast US has been developed and applied in clinical practice, and its assistance in the morphological analysis of breast masses has improved diagnostic accuracy and sensitivity.[ A recently developed, commercially available DL-CAD software for breast US (S-Detect; Samsung Medison Co., Seongnam, Korea) provides computer-based analysis of breast tumors based on morphologic features using a novel feature extraction technique and support vector machine classifier that provides a dichotomized final assessment of breast mass, possibly benign or possibly malignant, based on the American College of Radiology Breast Imaging Reporting and Data System ultrasonographic descriptors.[

In addition to lesion differentiation, we assume that the quantitative morphological information regarding breast cancer obtained on US would provide histopathologic information, including molecular subtypes, similar to quantitative radiomics analysis using MRI.[ Previous studies have revealed that irregular shape and spiculated or indistinct margins with posterior acoustic shadowing are associated with the luminal subtype, and oval to round shape and circumscribed margins with posterior acoustic enhancement are common in triple-negative breast cancer (TNBC) on breast US.[ However, the mass characteristics were assessed by qualitatively by radiologist, and quantitative assessment was not performed. The limitations of ultrasound, especially operator dependency, could be reduced by proving an association between quantitative assessment of ultrasound images using DL-CAD and histopathologic features, such as a specific tumor subtype.

Therefore, the purpose of our study was to investigate the correlations between ultrasonographic morphological characteristics, quantitatively assessed using DL-CAD, and histopathologic features of breast cancer.

Methods

Patients

We present the following article in accordance with the STROBE reporting checklist. The Institutional Review Board of Yeungnam University Hospital (IRB No. 2018-07-016) approved this retrospective study and waived the requirement for informed consent. All the methods in the study involving human participants were performed in accordance with relevant guidelines and regulations of the institutional and/or national research committee and with the 1964 Helsinki declaration and its lateral amendments or comparable ethical standards. Between February 2016 and April 2017, 459 women with newly diagnosed invasive breast cancer (on percutaneous biopsy) who had previously undergone preoperative breast US and subsequent surgery were included through a review of medical records at our institution, a tertiary academic hospital. Women were excluded if they had only ductal carcinoma in situ without invasive cancer on the surgical specimen (n = 84), the tumor was >5 cm (n = 32), had previously undergone neoadjuvant chemotherapy before surgery (n = 52), had previously undergone vacuum-assisted or excisional biopsy for diagnosis before preoperative US (n = 8), and had insufficient immunohistochemical (IHC) results on pathologic reports (n = 1). Only the tumors with the largest dimensions were included in women with multifocal or multicentric breast cancers. Finally, a total of 282 consecutive women (mean age, 54.4 years; range, 29–85 years) were included in our study.

Ultrasound image acquisition

Preoperative breast US was performed by one of the two board-certified breast radiologists (KYS and HMS with 5 and 25 years of experience in breast imaging, respectively) using iU22 (Philips Medical Systems, Bothell, WA) with a 5 to 12 MHz linear array transducer. All breast ultrasound examinations were performed in real-time with a handheld ultrasound probe, and bilateral whole-breast scanning was conducted. A standardized scanning protocol was used for every examination, using the transverse and sagittal orientations, with the inner aspect of the breast scanned with the patient in a supine position, and the outer aspect in supine oblique position, with the patient's ipsilateral arm raised above the head. The axilla was routinely scanned before the breast in our protocol.

Two additional board-certified breast radiologists (KYS and CJM with 5 and 13 years of experience in breast imaging, respectively) retrospectively reviewed the US images and selected the most representative image with consensus for each tumor for CAD analysis. Following image selection, each image was stored in the DICOM format.

Pathologic analysis

All patients underwent breast surgery, including breast-conserving surgery (n = 202, 71.6%) or mastectomy (n = 80, 28.4%). We reviewed each patient's pathological report of the surgical specimen to identify the invasive tumor size, histologic grade based on the Elston–Ellis system,[ histologic types of invasive cancer, lymph node (LN) status, and IHC analysis findings (i.e., ER, PR, and HER2 status, Ki67). For the interpretation of IHC analysis, semi-quantitative scorings of the percentage of positive cells with nuclear staining (range, 0–100%) was used for ER and PR expression levels. The cutoff value for defining ER and PR positivity was 1%.[ IHC analysis results were initially used to define HER2 expression, and tumors with a score of 3+ were defined as HER2-positive, and those with a score of 0 or 1+ were defined as HER2-negative. For cases with a score of 2+ on IHC, silver-enhanced in situ hybridization for the HER2 gene was performed to define HER2 expression. The percentage of the total number of tumor cells with nuclear staining was used to define the Ki67 index.[ Breast cancers were divided into 4 subtypes based on the IHC results: luminal A (ER- and/or PR-positive, HER2-negative, Ki67 ≤ 20%), luminal B (ER- and/or PR-positive, either HER2-positive or HER2-negative with Ki67 > 20%), HER2-positive (HER2-positive, ER- and PR-negative), and triple-negative (ER-, PR-, and HER2-negative).[

Analysis of quantitative morphologic scores using DL-CAD

A commercially available DL-CAD software (S-Detect; Samsung Medison Co., Seongnam, Korea) for breast US was used to obtain quantitative morphological information on mass features, and the final assessment was performed at a dedicated workstation. It provides computer-based analysis of tumor morphology using a novel feature extraction technique and a support vector machine classifier.[ The current commercially available version of the DL-CAD software only displays the final assessments in a dichotomized form as “possibly benign” or “possibly malignant” based on the maximum value of each breast imaging-reporting and data system (BI-RADS) descriptor (shape, margin, orientation, echo pattern, and posterior echogenic features), and does not present the quantitative morphologic scores. However, with technical support from Samsung Medison Co. (Seongnam, Korea), we used the original outputs (quantitative scores) of the DL-CAD software to analyze the US images. In the algorithm used, deep learning technology was applied during the generation of quantitative morphologic scores to build a classifier for BI-RADS descriptors. A series of layers of simple components constitute a network, each having their own nonlinear mappings between the input and output. In contrast to conventional machine learning, where human experts need to select representative imaging features, deep learning algorithms do not require human input. Instead, they determine the manner in which internal parameters are the best representations of data from large high-dimensional datasets via learning procedures.[ Combining the output for each BI-RADS descriptor with that of the other network for classifying the region of interest (ROI) images results in the final decision.[

Using DL-CAD software, the radiologists indicated the center of the mass, and a ROI was automatically drawn along the border of the mass. When the mass boundary was inadequately drawn, manual correction was performed. It automatically generated quantitative output values in a range between 0 and 1 for mass shape, orientation, margin, posterior features, and echo pattern for ROI-based masses on US according to the 5th edition of the BI-RADS lexicon (Fig. 1). To simplify the analysis, the internal values of shape, orientation, margin, and echo pattern were collected and arranged in a dichotomized manner (i.e., shape: irregular vs not-irregular, margin: circumscribed vs not-circumscribed, echogenicity: hypoechoic vs not-hypoechoic, orientation: parallel vs not-parallel).

Figure 1

Analysis of morphologic score of tumor on US using deep learning based computer-aided diagnosis (DL-CAD) software. A region of interest (ROI) (red line) was automatically drawn along the border of the tumor in a 46-year-old woman, and the US morphologic features were analyzed by the DL-CAD based on the ROI. The table shows quantitative scores of this tumor obtained from DL-CAD software (∗ pectoralis muscle under the lesion).

Data and statistical analysis

The quantitative morphological scores obtained using the DL-CAD software and the histopathological findings of the subsequent surgical specimens of 282 tumors were reviewed. We analyzed the correlation between the pathological features (mass size, tumor histologic grade, LN status, molecular subtype, etc) and quantitative scores of masses for each lexicon on US. The tumors were subdivided into 3 groups according to the size of the mass: <1 cm, 1 to 2 cm, and >2 cm. The quantitative scores of the masses on US were analyzed according to histopathological features using Student t test or one-way analysis of variance. In addition, multiple linear regression analysis with the stepwise selection method was used to determine the relative influence of the different histopathological features on the quantitative scores of masses for each lexicon on US. All statistical analyses were performed using SPSS statistics version 25 for Windows (SPSS Inc., Chicago, IL), and a P value <.05 was considered to indicate a significant difference.

Results

Baseline characteristics

A total of 282 tumors from 282 consecutive women were included in our study. The mean size of the 282 invasive tumors was 1.8 ± 1.0 (standard deviation) cm. For the histologic type of breast cancer, invasive ductal carcinoma, not otherwise specified (IDC, NOS) was the most common histologic type (n = 236, 83.7%), followed by invasive lobular carcinoma (ILC) (n = 18, 6.4%). There were 192 (68.1%) pathologic T1-stage tumors and 90 (31.9%) T2-stage tumors. Most cases were LN-negative (n = 174, 61.7%). The IHC results showed that luminal A tumors were the most common tumors (n = 144, 51.1%), followed by luminal B tumors (n = 77, 27.3%), TNBC (n = 39, 13.8%), and HER2-positive tumors (n = 22, 7.8%). Patient and tumor characteristics are listed in Table 1.

Table 1

Patient and tumor characteristics.

Characteristics	Value
Patient characteristics (n = 282)
Age, y
Mean ± standard deviation	54.4 ± 11
Median, range	54, 29–85
Operation
Breast-conserving surgery	202 (71.6)
Mastectomy	80 (28.4)
Tumor characteristics (n = 282)
Invasive tumor size, cm
<1	41 (14.5)
1–2	151 (53.6)
>2	90 (31.9)
Histologic type
Ductal, NOS	236 (83.7)
Lobular	18 (6.4)
Others	28 (9.9)
Histologic grade
I	53 (18.8)
II	90 (31.9)
III	139 (49.3)
Lymph node status
Negative	174 (61.7)
Positive	108 (38.3)
ER
Negative	62 (22.0)
Positive	220 (78.0)
PR
Negative	90 (31.9)
Positive	192 (68.1)
HER2
Negative	240 (85.1)
Positive	42 (14.9)
Subtype
Luminal A	144 (51.1)
Luminal B	77 (27.3)
HER2-positive	22 (7.8)
Triple-negative	39 (13.8)

US morphological characteristics associated with histopathologic features

We analyzed the correlations between US quantitative morphological scores obtained using DL-CAD and pathologic features, including tumor size, tumor histologic grade, LN status, receptor status, and tumor subtype. In the univariate analysis, the irregular shape score on US was higher in tumors with specific pathologic characteristics, including larger invasive size, positive LN status, and ER-positive and luminal A subtypes. Not-circumscribed margin (P < .001) and hypoechogenicity (P = .003) scores on US correlated with pathologic tumor size, and not-parallel orientation score on US correlated with histologic grade (P = .024) (Table 2). The quantitative scores for margin and echogenicity on US did not differ according to certain pathologic characteristics, including histologic grade, LN status, ER, PR, HER2 status, and tumor subtype. Multiple linear regression analysis revealed that pathologic tumor size was the only significant independent factor associated with quantitative US scores for irregular shape (P < .001) and not-circumscribed margin (P < .001).

Table 2

Correlations between histopathologic features and results of quantitative analysis of invasive breast cancer using DL-CAD.

	Shape		Margin		Echogenicity		Orientation
Pathologic feature	Irregular	P-value	Not-circumscribed	P-value	Hypoechoic	P-value	Not-parallel	P-value
Invasive tumor size, cm		<.001		<.001		.003		.506
<1.0	0.32 ± 0.33		0.49 ± 0.43		0.70 ± 0.31		0.39 ± 0.41
1.0–2.0	0.58 ± 0.31		0.77 ± 0.29		0.82 ± 0.26		0.47 ± 0.38
>2.0	0.74 ± 0.29		0.88 ± 0.25		0.87 ± 0.21		0.44 ± 0.35
Histologic grade		.529		.404		.054		.024
I (n = 53)	0.56 ± 0.34		0.72 ± 0.35		0.74 ± 0.32		0.53 ± 0.39
II (n = 90)	0.62 ± 0.34		0.80 ± 0.30		0.82 ± 0.26		0.49 ± 0.39
III (n = 139)	0.59 ± 0.33		0.76 ± 0.34		0.84 ± 0.24		0.39 ± 0.36
LN status		.001		.097		.146		.939
Negative (n = 174)	0.55 ± 0.34		0.74 ± 0.34		0.80 ± 0.27		0.45 ± 0.39
Positive (n = 108)	0.66 ± 0.31		0.81 ± 0.31		0.84 ± 0.24		0.45 ± 0.36
ER status		.016		.283		.164		.004
Negative (n = 62)	0.50 ± 0.34		0.73 ± 0.33		0.86 ± 0.23		0.33 ± 0.34
Positive (n = 220)	0.62 ± 0.33		0.78 ± 0.33		0.81 ± 0.27		0.48 ± 0.38
PR status		.249		.374		.491		.021
Negative (n = 90)	0.56 ± 0.34		0.74 ± 0.34		0.83 ± 0.24		0.37 ± 0.34
Positive (n = 192)	0.61 ± 0.33		0.78 ± 0.32		0.81 ± 0.27		0.48 ± 0.39
HER2 status		.171		.617		.522		.007
Negative (n = 240)	0.61 ± 0.33		0.77 ± 0.33		0.82 ± 0.26		0.47 ± 0.38
Positive (n = 42)	0.53 ± 0.33		0.74 ± 0.34		0.79 ± 0.28		0.31 ± 0.32
Subtype		.048		.296		.587		.002
Luminal A (n = 144)	0.63 ± 0.32		0.80 ± 0.31		0.80 ± 0.27		0.53 ± 0.38
Luminal B (n = 77)	0.60 ± 0.33		0.73 ± 0.36		0.82 ± 0.25		0.39 ± 0.36
HER2 (n = 22)	0.55 ± 0.30		0.77 ± 0.29		0.85 ± 0.24		0.36 ± 0.32
TNBC (n = 39)	0.47 ± 0.36		0.70 ± 0.35		0.86 ± 0.24		0.31 ± 0.35

US morphological characteristics associated with pathologic molecular subtypes

Since the pathologic size of the mass was the strongest factor associated with morphological characteristics on US, we analyzed the correlation between the quantitative score obtained using DL-CAD on US and the pathologic molecular subtype in T1-stage breast cancer to reduce the influence of tumor size on the analysis. The results of the multiple linear regression analyses are presented in Table 3. The pathologic molecular subtype independently correlated with the quantitative US scores of the irregular shape, not-circumscribed margin, and not-parallel orientation. Luminal A tumors (Fig. 2) showed higher US scores for irregular shape than TNBC (Fig. 3) and higher US scores for not-circumscribed margin and not-parallel orientation than luminal B tumors or TNBC. However, for tumors >2 cm in size, no significant variables were observed. As the pathologic size of the invasive tumor increased, the tumor tended to show a more irregular shape without a circumscribed margin on the US, regardless of the pathologic molecular subtype (Fig. 4).

Table 3

Multiple linear regression analysis of molecular subtypes of breast cancer associated with quantitative scores in T1 breast cancer.

Variable	B	Standard error	P-value
Irregular shape score
Subtype
Lum A	…	…	…
Lum B	−0.110	0.057	.055
HER2	−0.088	0.088	.316
TNBC	−0.307	0.074	<.001
Not-circumscribed margin score
Subtype
Lum A	…	…	…
Lum B	−0.145	0.061	.018
HER2	−0.019	0.094	.840
TNBC	−0.182	0.080	.024
Not-parallel orientation score
Subtype
Lum A	…	…	…
Lum B	−0.250	0.066	<.001
HER2	−0.170	0.102	.096
TNBC	−0.302	0.087	.001

Figure 2

Breast ultrasound image of a 67-year-old woman showing an 11-mm luminal A subtype, grade II invasive carcinoma, not otherwise specified. On the grayscale ultrasound image, the region of interest is drawn (in red color) along the border of the mass using the deep learning-based computer-aided diagnostic system (A). The raw data for quantitative scores are irregular shape, 0.0915; not-circumscribed margin, 0.637; hypoechoic echogenicity, 0.959; and not-parallel orientation, 0.971. The gross photograph of tumor specimen (white arrow, B) and low-power hematoxylin and eosin (H&E, ×7) slide of this tumor (C) revealed the irregular shape and not-circumscribed margin of this tumor (∗ pectoralis muscle under the lesion).

Figure 3

Breast ultrasound image of a 47-year-old woman with a 15-mm TNBC subtype, grade III invasive carcinoma, not otherwise specified NOS. On the grayscale ultrasound image, the region of interest is drawn (in red color) along the border of the mass using the deep learning-based computer-aided diagnostic system by DL-CAD (A). The raw data for quantitative scores are irregular shape, 0.020; not- circumscribed margin, 0.019; hypoechoic echogenicity, 0.990; not-parallel orientation, 0.004. The gross photograph of tumor specimen (white arrow, B) and low-power hematoxylin and eosin (H&E, ×7) slide of this tumor (C) revealed the oval shape and circumscribed margin of this tumor (∗ pectoralis muscle under the lesion). DL-CAD = deep learning-based computer-aided diagnosis, NOS = not otherwise specified.

Figure 4

Breast ultrasound image of a 73-year-old woman with a 27-mm TNBC subtype, grade III invasive carcinoma, not otherwise specified NOS. On the grayscale ultrasound image, the region of interest is drawn (in red color) along the border of the mass by using the deep learning-based computer-aided diagnostic system DL-CAD. The raw data for quantitative scores are irregular shape, 0.991; not-circumscribed margin, 0.998; hypoechoic echogenicity, 0.957; not-parallel orientation, 0.423. The gross photograph of tumor specimen (white arrows, B) and low-power hematoxylin and eosin (H&E, ×7) slide of this tumor (C) revealed the irregular shape and not-circumscribed margin of this tumor. Since the large size of the tumor, the entire morphology of tumor was not included in the one same slide (∗ pectoralis muscle under the lesion). DL-CAD = deep learning-based computer-aided diagnosis, NOS = not otherwise specified.

Discussion

Our study demonstrated that quantitative morphological scores obtained using DL-CAD with B-mode breast US correlated with certain pathologic tumor characteristics, including tumor size, histologic grade, LN status, and receptor status. Among T1 stage breast cancers, luminal A tumors exhibited more irregular features with no parallel orientation on US, whereas TNBC showed rounder/more oval and parallel orientation on US.

In current clinical practice, US is a widely used non-invasive medical imaging technique for breast cancer. Breast cancer is a highly heterogeneous disease and the tumor subtype determined by IHC analysis is critical for determining the treatment options and prognosis.[ However, IHC analysis has certain limitations. Owing to tumor heterogeneity, the sampling and analysis of the tumor tissue are uncertain, and visual interpretations are subjective and may lead to misinterpretations.[ Radiomics is the study identifying the relationship between tumor characteristics at the cellular or genetic level and morphologic characteristics on medical images.[ It is hypothesized that comprehensive features of the entire tumor on medical imaging could reveal predictive associations between the images and medical outcomes.[ Breast cancer has been the primary focus of radiomics research, and in these studies, the luminal subtype tumor commonly presented as a mass with a poorly circumscribed margin on MG and US and showed posterior acoustic shadowing on US.[ HER2-positive subtype tumors are often accompanied by calcifications on MG and commonly present as irregularly shaped masses on US, and washout or fast initial kinetics on MRI.[ TNBC is often observed as a non-calcified, relatively circumscribed mass on the MG and a circumscribed mass with posterior acoustic enhancement on US.[ Quantitative evaluation using whole-tumor histogram-based imaging features derived from apparent diffusion coefficient maps and dynamic contrast-enhanced (DCE) MR semi-quantitative maps or multiparametric MRI have provided useful information for differentiating TNBC from other subtypes on MRI.[

Thus far, radiomics research using US has been limited owing to its operator dependency and subjective interpretation characteristics. The recent development of DL-CAD enabled the morphological analysis of breast mass based on the raw data of quantitative scores for the BI-RADS lexicon. Several studies have reported that DL-CAD can help improve the diagnostic performance, especially accuracy and specificity, of breast US for distinguishing benign from malignant lesions.[ However, to the best of our knowledge, no study has analyzed raw data itself, driven by DL-CAD, to quantify the morphological characteristics of breast cancers. Kim et al[ analyzed the diagnostic performance of breast US using quantitative variables, but the quantitative variables in that study were width, height, height/width ratio, area, and depth. This is the first study focusing on the correlation between molecular subtype and sonographic features of breast cancer using quantitative analysis by DL-CAD with breast US.

Our DL-CAD-based results are consistent with those of previous studies revealing that luminal subtype tumors tend to present as a mass with a poorly circumscribed margin, HER2-positive subtype as an irregular mass with a not-circumscribed margin, and TNBC as a distinct mass with a circumscribed margin on US when the tumor size is <2 cm.[ However, an interesting finding was that as the pathologic size of the invasive tumor increased, the tumor tended to show a more irregular shape without a circumscribed margin on the US regardless of the molecular subtype, which meant that the molecular subtype of the tumor was significantly correlated with the ultrasonographic morphological characteristics only when the tumor size was below the T1 stage. Thus, the imaging phenotypes of breast cancer should be cautiously interpreted based on tumor size.

Our study had some limitations. First, this was a retrospective study from a single center with a limited number of cases. We used representative still images stored in the PACS system instead of real-time ultrasonographic examinations during image analysis by DL-CAD; thus, there is the possibility of selection bias and reader variability in determining the representative images of breast cancer. Second, non-mass lesions were not included in this study. Third, our findings do not suggest a clinically useful cutoff value by DL-CAD to assign tumor molecular subtypes using US images.

In conclusion, using DL-CAD, we demonstrated that quantitative analysis of the morphological characteristics of breast cancers on US correlated with the histopathologic features and could provide useful information regarding the imaging phenotypes of breast cancer.

Acknowledgments

Samsung Medison Co. (Seongnam, Korea) provided technical support for the US image analysis with DL-CAD software (S-Detect; Samsung Medison Co., Seongnam, Korea) and obtained outputs from the algorithm.

Author contributions

Conceptualization: Jung Min Chang.

Data curation: Young Seon Kim, Seung Eun Lee, Soo-Yeon Kim, Young Kyung Bae.

Formal analysis: Young Seon Kim, Seung Eun Lee, Jung Min Chang, Soo-Yeon Kim, Young Kyung Bae.

Funding acquisition: Jung Min Chang, Young Seon Kim.

Investigation: Young Seon Kim, Seung Eun Lee, Soo-Yeon Kim.

Methodology: Young Seon Kim.

Supervision: Jung Min Chang.

Visualization: Young Seon Kim, Young Kyung Bae.

Writing – original draft: Young Seon Kim.

Writing – review & editing: Young Seon Kim, Jung Min Chang.