By retrospectively reviewing two tertiary medical centers’ database (October 2013 and October 2020), we identified 948 clinically T1-2 N0 patients who had negative AUS, subsequent breast MRI, and BCT. At first, a primary development set of 402 patients from Korea University Anam Hospital were retrospectively analyzed. Inclusion criteria were having preoperative breast and AUS by experienced breast radiologists prior to breast MRI and negative axillary result on ultrasound. Exclusion criteria were positive axillary result on ultrasound, ultrasound performed by an outside hospital/resident, NAC before operation, distant metastasis at diagnosis, inadequate information about immunohistochemical staining, and outside MRI. A validation set of 546 patients from Korea University Guro Hospital was further analyzed. The diagram for establishing the development and validation sets is summarized in Fig.1.

To evaluate the clinical T and N stage, one of three breast radiologists (K.R.C., S.E.S., E.K.P., with 21,11, and 8 years of experience in breast imaging, respectively) in Korea University Anam Hospital and one of two breast radiologists (O.H.W. and H.S., with 19 and 7 years of experience) in Korea University Guro Hospital performed preoperative breast and AUS with a Logiq 9 unit (General Electric, Milwaukee, USA), and an Aixplorer system (SuperSonic Imagine, Aix en Provence, France) within a median of 14 and 12 days before surgery in the two sets. All cancers were assessed using the breast imaging reporting and data system (BI-RADS) categories 4A, 4B, 4C, and 5 (Fig. 2A). Axillary results on AUS were recorded as either negative or positive (Fig. 2B). Based on BI-RADS, positive axilla was considered when a LN with at least one of the following features was found: eccentric cortical thickening, diffuse cortical thickening of 3 mm or large, rounded hypoechoic, complete or partial effacement of the fatty hilum, and nonhilar cortical blood flow on color Doppler images.2123 If the assessment for axillary LN were equivocal or borderline, two breast radiologists (K.R.C., S.E.S. with 21 and 11 years of experience in breast imaging, respectively) re-evaluated the axillary LN status and categorized it into negative or positive by consensus. For obtaining clinical T stage, the longest tumor size on ultrasound was measured.

Breast MRI examinations were performed using two 3.0T scanners (Achieva 3.0T TX; Philips Healthcare, Best, the Netherlands and MAGNETOM Skyra; Siemens Healthcare, Erlangen, Germany) in Korea University Anam Hospital and a 3.0T scanner (MAGNETOM Skyra; Siemens Healthcare) in Korea University Guro Hospital with a dedicated breast array coil in the axial orientation. All MRI scans were taken with the patients in a prone position. Detailed protocols were described in Supplementary Data 1.

One of three breast radiologists in Korea University Anam Hospital and one of two breast radiologists in Korea University Guro Hospital read breast MRI. Axillary MRI results were recorded as negative or positive (Fig. 2C). Positive axilla on MRI was considered when a LN with at least one of the following features was observed: markedly enlarged and morphologically grossly abnormal LNs, especially when they are distinctly different from other visible LNs, cortical irregularity or apparent spiculation, loss of fatty hilum, round shape, or a long axis to short axis ratio of less than two.2324 Axillary LN with equivocal or borderline assessment at radiologic report were re-evaluated in same way as breast ultrasound.

The mammographic and MRI features were retrospectively analyzed in a blinded manner without clinical information by two breast radiologists (K.R.C., S.E.S., with 21 and 11 years of experience in breast imaging, respectively) using the 2013 BI-RADS lexicon.25 Mammographic findings were evaluated with regard to the calcifications (presence vs. absence), and tumor location (outer vs. inner) (Fig. 2D). The shape, margin, density and echogenicity of tumors on mammography and ultrasound were not evaluated because we intended to use those findings from MRI. MRI features were evaluated in terms of mass shape (round to oval vs. irregular), mass margin (circumscribed or irregular vs. spiculated), mass internal enhancement (homogeneous or heterogeneous vs. rim), intratumoral high signal intensity and peritumoral edema on T2-weighted image (presence vs. absence) (Fig. 2E).

Clinical data (menopause, symptom, family history of breast cancer) were reviewed from the electronic medical records. Imaging data were obtained from radiology reports. Pathological data were acquired from histopathological reports of surgery. Pathologic data of primary breast cancer and LN metastasis were evaluated from surgical and pathologic reports. According to the eighth edition of the American Joint Committee on Cancer (AJCC) staging, pN1mi which means micrometastases (larger than 0.2 mm, but none larger than 2.0 mm) at LN was defined as pN1.21 At least one LN metastasis including pN1mi was considered as disease positive. In addition, HNB was defined as ≥ 3 positive LNs and low nodal burden was defined as 1–2 positive LNs according to the Z0011 trial.45 Estrogen receptor (ER) and progesterone receptor (PR) positivity were defined as the presence of positive staining in at least 1% of the nuclei in ten high-power fields. Human epidermal growth factor receptor 2 (HER2) negativity was defined as an immunohistochemical score of 1+ or 2+ staining with negative HER2 gene amplification on fluorescence in situ hybridization. Histologic type was categorized as ductal or other, and histologic grade was categorized as low (grade 1 or 2) or high (grade 3). Molecular subtypes were categorized as luminal (ER and/or PR-positive and HER2-negative), HER2-enriched (ER/PR-positive or negative, HER2-positive), or triple-negative (ER-negative, PR-negative, and HER2-negative).22 The Ki-67 proliferation index was dichotomized as low (< 14%) or high (≥ 14%).22

Baseline characteristics were compared using the Mann-Whitney U test, independent t-test, or independent χ² test. The diagnostic performance of breast MRI and nomogram was assessed using sensitivity, specificity, PPV, and NPV. Logistic regression analysis was performed to identify clinical and imaging factors associated with axillary LN metastasis. For multivariable analysis, covariates with P values < 0.05 in the univariable analysis with forward elimination were used. Based on the adjusted odds ratio (OR) from multivariate analysis, clinical-imaging model was constructed and the nomogram was developed to visualize the results of the predictive model. Internal and external validation of nomogram was performed in development and validation sets. For validation, discrimination and calibration were assessed using area under the curve (AUC) and calibration slope. All statistical analyses were performed by using SPSS for Windows (version 20.0; IBM Corp., Armonk, NY, USA) and open-source R software (version 3.5.1; R Foundation for Statistical Computing, R Foundation, Vienna, Austria).

The Institutional Review Board (IRB) of Korea University Anam Hospital approved this retrospective study (IRB No.: 2021AN0464). The IRB waived the need for informed consent.

A total of 948 consecutive female patients were included (mean age ± standard deviation, 55.8 ± 10.9; range, 28–87 years) with 402 female patients (57.61 ± 11.58) in the development set and 546 patients (54.43 ± 10.02) in the validation set. A comparison of characteristics between the two sets is listed in Table 1. Compared to the development set, the validation set was composed of younger patients, pathologic T2 stage, lymphovascular invasion, luminal subtype, higher Ki-67, BI-RADS 5 assessment category on US, and positive axilla on MRI (all P < 0.05).

The FNRs of AUS were 20.1% (81 of 402 patients) within development set and 19.2% (105 of 546 patients) within validation set. The rates of HNB were 1.2% (5/402) and 2.2% (12/546), respectively. According to the clinical T stage, the rates of HNB were 0.6% (2/337) in T1 stage and 4.6% (3/65) in T2 stage within development set, and 2.5% (10/403) in the T1 stage and 1.4% (2/143) in the T2 stage within validation set.

To investigate the factors that can predict HNB, multivariable analysis was performed (Supplementary Table 1). Interestingly, positive axilla on MRI showed marginal significance for predicting HNB (adjusted OR, 2.7; P = 0.059), following after lymphovascular invasion (adjusted OR 3.4; P = 0.023).

FNR of breast MRI were 16.3% (60 of 367 patients) within development set and 13.4% (64 of 476 patients) within validation set. For predicting additional LN metastasis in patients with negative AUS, the NPVs of breast MRI in both development and validation sets were high (83.6% [307/367] and 86.5% [412/476]) while PPVs were low (60.0% [21/35] and 58.5% [41/70]). The sensitivities were low (25.9% [21/81] and 39.0% [41/105]), whereas the specificities were high (95.6% [307/321] and 93.4% [412/441]) (Table 2).

Among clinical factors, lymphovascular invasion, higher pathologic T stage, and PR positivity were factors associated with LN metastasis at univariable analysis (Table 3). Among imaging factors, outer tumor location on mammography, higher clinical T stage measured on US, BI-RADS 5 assessment category on ultrasound, and positive axilla on MRI were significant factors (Table 4). For development of clinical-imaging nomogram which could be used preoperatively, we selected significant clinical and imaging predictors which could be obtained before operation. Therefore, pathologic T stage and lymphovascular invasion which could be assessed after operation were not selected. At multivariable analysis, significant clinical-imaging predictors were PR positivity (adjusted OR, 2.8; P = 0.017), outer tumor location on mammography (adjusted OR, 1.8; P = 0.043), BI-RADS category 5 on US (adjusted OR, 2.6; P = 0.002), and positive axilla on MRI (adjusted OR, 7.8; P < 0.001). However, clinical T2 stage measured on ultrasound did not acquire the significance (adjusted OR, 1.5; P = 0.260) (Table 5). The nomogram integrating clinical and imaging factors was constructed (Fig. 3). The AUC of the clinical-imaging nomogram was 0.76 (95% confidence interval [CI], 0.70–0.81) in the development set and 0.77 (95% CI, 0.70–0.81) in the validation set (Fig. 4). The NPV of clinical-imaging nomogram in the validation set was 82.0%.