Radiomics Model Based on Shear-Wave Elastography Can Predict the Axillary Lymph Node Status in Early-Stage Breast Cancer

Background: Accurate prediction of axillary lymph node (ALN) involvement in early-stage breast cancer is important for determining appropriate axillary treatment and therefore avoiding unnecessary axillary surgery and complications. This study aimed to develop and validate an ultrasound radiomics nomogram for preoperative evaluation of the ALN burden. Methods: Data of 303 patients from Wuhan Tongji Hospital (training cohort) and 130 cases from Hunan Provincial Tumour Hospital (external validation cohort) between Jun 2016 and May 2019 were enrolled. Radiomic features were extracted from shear-wave elastography (SWE) and corresponding B-mode ultrasound (BMUS) images. Then, the minimum redundancy maximum relevance (MRMR) and least absolute shrinkage and selection operator (LASSO) algorithm were used to select ALN status-related features and construct the SWE and BMUS radiomic signatures. Proportional odds ordinal logistic regression was performed using the radiomic signature together with clinical data, and an ordinal nomogram was subsequently developed. We evaluated the performance of the nomogram using C-index, calibration, and compared it with clinical model. Results: Multivariate analysis indicated that SWE signature, US-reported LN status and molecular subtype were independent risk factors associated with ALN status. The radiomics nomogram based on these variables showed good calibration and discrimination in the training set (overall C-index: 0.842; 95%CI, 0.773–0.879) and the validation set (overall C-index: 0.822; 95%CI, 0.765–0.838). For discriminating between disease-free axilla (N0) and any axillary metastasis (N + ( ≥ 1)), it achieved C-index of 0.845 (95%CI, 0.777–0.914) for the training cohort and 0.817 (95%CI, 0.769–0.865) for the validation cohort. The tool could also discriminate between low (N + (1–2)) and heavy metastatic burden of ALN (N + ( ≥ 3)), with C-index of 0.827 (95%CI, 0.742–0.913) for the training cohort and 0.810 (95%CI, 0.755–0.864) for the validation cohort. Conclusions: The presented radiomics nomogram shows favourable predictive ability for ALN staging in patients with early-stage breast cancer, which could provide incremental information for preoperative decision-making.

compared with ALND, but the incidence of potentially serious allergic reaction due to blue dye [6], radiation and costs related to radiocolloid injection [7], could not be eliminated. Besides, SLN biopsy and intraoperative frozen section examination would increase the operating and anaesthetic time.
The American College of Surgeons Oncology Group Z0011 (ACOSOG Z0011) trial revealed that among women with clinical T1/T2 breast cancer, if there were 1 or 2 SLNs involved, the spared of an ALND would not lead to inferior overall and disease-free survival [4,5]. Thus, the number of involved ALNs may have great impact on patients' long-term survival outcome. According to the number of positive metastatic ALNs, we can categorize the patients into three subgroups: disease-free axilla (N0), low metastatic burden of ALN (N + (1-2)), and heavy metastatic burden of ALN (N + ( ≥ 3)), which is important for determining the extent of axillary treatment [4,8,9]. There were studies demonstrating that about 43-65% of patients who had positive SLNs underwent unnecessary ALND because of no additional non-SLN metastasis was detected, leading to high axillary morbidity [10,11]. If there was reliable none-invasive preoperative methods for predicting different ALN burden, then individualized and precise minimally invasive treatment could be achieved to reduce the unnecessary SLND or ALND [12].
Two-dimensional (2D) shear wave elastography (SWE) is an elastographic technique that integrates Bmodel ultrasound (BMUS) with a color-coded map to allow for better characterization of breast lesions [13], which showed promise in distinguishing malignant and benign breast tumour [14]. Previous studies revealed that stiffness of breast cancer was also a predictor of ALN status, for higher shear wave velocity of breast cancer was related to higher possibility of ALN metastasis [15,16]. However, traditional SWE parameters can only capture limited characteristics of the heterogeneous tumour, the area under the receiver operating characteristic curves (AUCs) for predicting ALN metastasis were 0.585-0.719 [17], which is unsatisfactory for clinical application.
Radiomics is an emerging technique that converts medical images into high-throughput features, which is valuable for tumour phenotyping [18]. Nomogram that combined radiomic signature extracted from US or MRI and clinicopathological factors has shown potential on discriminating N0 and any axillary metastasis (N+) [19,20]. However, the use of radiomics nomogram to discriminate disease-free axilla, low ALN burden and heavy ALN burden has yet to be reported.
To address this, we aimed to develop and validated an ordinal nomogram incorporating 2D SWE radiomic signature for evaluating ALN status in early stage breast cancer. We focused on preoperatively discriminate pathologic N0, N + (1)(2), and N + (≥3), because an accurate ALN burden evaluation is the basis of individual axillary treatment.

Study population
This retrospective study (clinical trial ChiCTR1900027676) was approved by the Institutional Review Board of the participant hospitals, and informed consent was waived. Between June 2016 and May 2019, consecutive patients with primary breast cancer in hospital #1 (Tongji Hospital of Huazhong University of Science and Technology, training cohort) and hospital #2 (Hunan Provincial Tumour Hospital, validation cohort) referred for 2D SWE examination with subsequent biopsy and surgical treatment were evaluated.
Patient's inclusion criteria including: (1) pathologically con rmed primary breast cancer; (2) all patients underwent mastectomy with SLND or ALND (in case SLN biopsy was positive); (3) SWE image of breast displayed with the same BMUS image in split-screen mode was carried out within two weeks prior to operation; and (4) availability of clinical data. The exclusion criteria including: (1) preoperative therapy (neoadjuvant radiotherapy or chemotherapy) before US examination; (2) with multifocal lesions or bilateral disease; (3) little or no shear wave signal was obtained in the region of interest (ROI) of SWE (masses deeper than 3 cm in depth will lead to the attenuation of SWE); (4)  from Hunan Provincial Tumour Hospital was enrolled with the same criteria. A owchart describing the patient screening process is shown in Figure 1.
The US-reported LN status was obtained from the US reports, and axillary images containing important features of suspicious LNs were documented into the Picture Archiving and Communication Systems (PACS). It was retrospectively reviewed and veri ed by two radiologists (X.M.L and S.C.T, with 15 and 31 years of experience respectively). US features of LN used to assess suspicion for malignancy were as follows: 1) irregular cortical thickness ≥3 mm; 2) longest/shortest axes ratio < 2; or 3) absence of fatty hilum [21].
The baseline clinicopathological data were derived from the patient medical records, including age, clinical tumour size, pathological type, IHC results and post-operative ALN status. According to the 2017 St Gallen International Expert Consensus, the breast cancers were classified into four molecular subtypes based on preoperative biopsy: human epidermal growth factor receptor-2 positive (HER2+), triplenegative, Luminal A, and Luminal B [22]. The status of HER2, progesterone receptor (PR), estrogenic receptor (ER) and Ki-67 was assessed by IHC examination.

US image acquisition
BMUS and SWE examinations were performed with a Supersonic Aixplorer system (SuperSonic Imagine, Aix-en-Provence, France) using a 4-14 MHz linear transducer by ve radiologists from Tongji Hospital and two radiologists from Hunan Tumour Hospital experienced in breast US according to standard protocols.
After standard conventional BMUS, SWE was performed. The ROI was set to include the whole breast cancer and adjacent normal parenchyma for SWE acquisition, and stiffness was shown as a colour map on the ROI. On the colour map, blue and red regions re ect comparatively soft (low kPa) and stiff (high kPa) tissues, respectively [23]. The detail US examination procedures are presented in Additional le 1.
Tumour segmentation and radiomic feature extraction Figure 2 shows the flowchart of the radiomics work ow. One image per tumour was used for analysis. The ROI for feature extraction was manually delineated on the largest cross section of B-model image using ITK-SNAP software (3.8.0; http://www.itksnap.org). All the manual segmentations were performed by two experienced breast US radiologists (with 11 and 9 years of experience, respectively) and each twice who were blinded from the final pathological diagnosis (for interobserver and intraobserver reproducibility evaluation). Because the contour of the lesion on SWE was indefinite, the same region on BMUS was copied and pasted to the corresponding SWE image, and was expanded to include the "stiff rim" sign if it existed. SWE is a combination of B-model image and pseudocolor elasticity layer, algorithms presented in previous studies were employed to produce clean quantitative images that mapping the tissue stiffness as grey levels [24][25][26]. The radiomic features were extracted automatically from each BMUS and SWE image by Pyradiomics (https://pyradiomics.readthedocs.io/en/latest/index.html) (Additional le 2, Supplemental Table 1 The stable features with interclass correlation coefficient (ICC) > 0.8 were selected to adapt different segmentations.

Radiomic signature building
We used Spearman's correlation coe cient to evaluate the relevance and redundancy of the features, and eliminated redundant features that with a Spearman's correlation coe cient ≥ 0.8. Then, the minimum redundancy maximum relevance (MRMR) algorithm and least absolute shrinkage and selection operator (LASSO) regression method using 10-fold cross-validation was applied to select the most useful predictive ALN status-related features from the training data set [28]. The formulas for the SWE and BMUS radiomic signatures were built using the respective selected features.

Radiomics nomogram construction
Univariate analysis was conducted to select statistically signi cant clinical factors associated with ALN metastasis. We mainly considered SWE radiomic signature in our nomogram; however, the incremental predictive value of BMUS radiomic signature to the model was also investigated using the net reclassification index (NRI) and integrated discriminatory improvement (IDI). Proportional odds ordinal logistic regression was used to build the radiomics nomogram based on the radiomic signatures and clinical characteristics [29]. The total points (Nomo-score) of each patients were calculated based on predictors of the nomogram. The association of the Nomo-score with pathologic ALN status was assessed using Spearman's correlation analysis. In addition, a classi cation procedure was proposed based on cut-offs of the Nomo-score to split patients into three subgroups of ALN status. Furthermore, a model based on clinical characteristics only was developed using the same method for comparison.

Performance evaluation
Harrell's C-indexes of the radiomics nomogram and clinical model were compared in the training and validation cohorts with Delong test. Besides, we carried out subgroup analysis based on patient age and clinical N stage, and calculated pairwise C-indexes for discriminating N + (≥1) versus N0, and N + (≥3) versus N + (1-2). The calibration curve was also plotted to measure the model. Among the subgroup analysis, N + (≥1) versus N0 is of special concern since it would determine the axillary surgical strategy.
Decision curve analysis (DCA) was conducted accordingly to evaluate the clinical usefulness of the model in guiding SLN biopsy by quantifying the net bene ts.

Statistical analysis
Statistical analyses were conducted with R software 3.6.1 and SPSS20.0 software (SPSS Inc., Chicago, IL). A two-sided P value less than 0.05 was used as the standard of statistically signi cant difference. In the univariate analysis, the differences in clinical characteristics between the patients of different groups were compared using Mann-Whitney U test or independent t test for continuous variables, and chi-square test or Fisher's exact test for categorical variables, as appropriate. Analysis of variance and Kruskal-Wallis H test were used for comparing more than two groups. The detailed descriptions of the LASSO and DCA algorithm are provided in the Additional le 1.

Patient characteristics
The baseline characteristics of patients and postoperative pathological information of breast lesions in the training and validation cohorts are displayed in Table 1 and Table 2. There were no signi cant differences between the two cohorts in age, clinical tumour size, status of hormone receptor (HR), HER2, ki67 and pathological type. According to the results of postoperative pathological analysis, 191 and 82 had disease-free axilla (N0), 46 and 26 had low ALN burden (N+(1-2)), 66 and 22 had heavy ALN burden (N+(≥3)) for the training and validation cohorts, respectively. There was no signi cant difference in ALN status between the two cohorts (P = 0.312). Of the total 433 patients, 262 (61%) had cT1 tumours and 171 (39%) had cT2 tumours. Among the 160 patients with ALN metastasis, 72 (45%) had one or two positive ALNs and 88 (55%) had three or more positive ALNs.

Feature selection and radiomics nomogram construction
Four-hundred two imaging features were extracted from each BMUS image. These features were reduced to six ALN status-related parameters after MRMR and LASSO algorithm in the training cohort (see Additional le 3, Supplemental Figure 1 a-b). Likewise, the 1558 SWE features were reduced to eight risk predictors by the same procedure (see Additional le 3, Supplemental Figure 1 c-d). Favourable interobserver and intraobserver reproducibility of feature extraction were achieved about these features, with intraobserver ICCs ranging from 0.803 to 0.891 and the interobserver ICCs ranging from 0.826 to 0.870. The BMUS and SWE radiomic signature calculation formulas are presented in the Additional le 1.
The multivariable regression analysis in the training cohort showed that SWE signature, molecular subtype and US-reported LN status were independent predictors for pathologic ALN status (Table 3). These predictors were combined into the radiomics nomogram (Figure 3a). The NRI and IDI analysis revealed that the addition of BMUS signature into the model did not show signi cantly better performance (NRI 0.0474, P = 0.130; IDI 0.0002, P = 0.861).

Model validation
As shown in Figure 3b, there was a signi cant positive correlation between Nomo-score and pathologic ALN status, which was also con rmed by Spearman's correlation coe cients (0.528 for training cohort and 0.580 for validation cohort, P < 0.001 for both, Figure 3c). As shown in Table 4 The calibration curves showed good agreement between the nomogram predicted outcomes and the real ALN status (Figure 4 a-b). If we use this model to guide lymphadenectomy (non-N0 patients receive SLN biopsy and N0 patients do not), as shown in Figure 4 c-d, the decision curves show that the radiomic nomogram could add more bene t to patients than the clinical model, traditional axillary US examination, non-SLN biopsy scheme, and all-SLN biopsy scheme.
The strati ed analysis showed that the performance of the radiomics nomogram was not affected by patient age, and was robust even in the US-reported LN negative subgroup (cN0) (Additional le 3, Supplemental Figure 2). Clinicians may be interested in how many patients with US-diagnosed N0 disease will be upstaged with the nomogram (non-N0 by pathology). These cases could be named as occult ALN metastasis, which are with no typical US signs. The results showed that the nomogram could well detect the patients with occult positive ALNs [82.1% (46/56) upgraded].

Discussion
In this study, we built an ordinal radiomics nomogram based on 2D SWE to predict the number of ALN metastases in breast cancer. The nomogram provided an easy-to-use and individualized tool for evaluating ALN status with high predictive ability, which can help tailor the optimum extent of axillary surgery. Traditionally, regional ALNs with features of enlarged size, irregular shapes, unclear margins, or loss of fatty on US imaging are suspicious for nodal involvement, and these were usually applied to guide fine-needle aspiration biopsy [21]. However, this standard showed relatively poor performance in our cohorts, for several patients (56/160) who had pathologically positive ALNs have no suspicious nodal signs on US imaging. In contrast, our nomogram performed signi cantly better than the routinely used axillary US-guided ALN detection method. Moreover, 82.1% of the occult ALN metastasis with no typical US signs (missed by the radiologists) were detected by the nomogram, which demonstrated that our model could be an important supplement to the current US examination.
According to ACOSOG Z0011 results, early-stage breast cancer with less than three SLN metastases had no inferior long-tern survival if they received SLND only rather than ALND [4,5]. Based on the ndings, patients should undergo SLND rst to clarify ALN status without considering whether they have clinically positive node or not [30]. However, SLND has false-negative rates ranging from 7.8-27.3%, which would lead to adverse outcomes including understaging the cancer and an increased risk of recurrence [31][32][33]. Our nomogram is able to distinguish patients with a negative ALN (N0) from those with any ALN status (N + (≥1)), and performed better than the routine axillary US examination. For patients with Nomo-score of ≤70, the percentage of heavy ALN burden was only 3.33% and 0% for the training and validation cohort respectively. Thus, surgeons may potentially forgo other examinations or exploratory surgery to confirm the absence of positive non-SLN metastasis based on the SLND and can thereby avoid excessive treatments. Moreover, in patients with Nomo-score that >100, about half of the patients showed heavy ALN burden, this will increase the surgeon's con dence to perform ALND based on a positive SLN biopsy. Therefore, our nomogram showed the promise for assisting decision-making for appropriate axillary treatment based on the current retrievable clinical information. Surgeons can then consider various factors, including the calculated Nomo-score of individual patients, and other preoperative clinical information, as well as their own clinical experience, to make a comprehensive judgment on the proper surgery plan.
Compared with previous studies that using SWE values to predict ALN status [15,17], our study obtained a better diagnostic performance by applying radiomic signature taken from the SWE images. Although the morphological and stiffness information of tumours can be easily discerned, high-dimensional radiomic features are still challenging to decipher through the naked eye (see Additional le 3, Supplemental Figure 3). Instead of measuring the breast cancer's stiffness based on the parameters of SWE, the whole shear wave ROI was quantified automatically into high throughput information and a large number of discriminative features could be acquired to phenotype the tumour. We further analysed the SWE radiomic features in the nomogram. Of note, the "wavelet-HL_gldm_DependenceNonUniformityNormalized_R" was the most signi cant radiomic feature of SWE in primary cancer that correlated with the ALN status, which measures the similarity of dependence throughout the image, with a higher value indicating more heterogeneity of the tumour. The other seven SWE radiomic features also re ect the heterogeneity of the tumour in different aspects, implied that the higher the heterogeneity of the primary cancer are, and thus the higher possibility of tumoral invasiveness and heavy burden of ALN metastases is.
Next, we considered the clinicopathological characteristics. St Gallen molecular subtype, which show the pervasive differences of breast cancer in their gene expression patterns [34], was found to be signi cantly associated with ALN status in this study. Van Limbergen and colleagues pointed out that the HER2enriched tumours were more likely to be lymph node positive [35]. Besides, Voduc and colleagues found that Luminal A tumours were associated with a lower risk of ALN metastasis than the other subtypes [36]. Our results have contributed to the accumulated evidence that molecular subtype could be a strong predictor of ALN status, and patients with positive HER2 were more inclined to ALN metastasis.
Our study has some unavoidable limitations. First, gene markers of breast cancer such as BRCA1 and BRCA2, are demonstrated helpful for patients risk strati cation [37]. However, a radio-genomics analysis that focused on the combination of genomics and radiomic phenotypes, is not available currently based on the retrospective datasets. Second, patients with multifocal and bilateral breast disease were excluded since it is hard to determine which lesion would responsible for the metastatic ALN and should be input in the radiomics model. Finally, since it was a retrospective study, some of the axillary US examinations may not have purposefully identified all abnormal appearing LNs. In previous literature [9], ≥3 abnormal LNs on US was reported to be a strong predictor of heavy nodal burden. The contribution of the exact number of suspicious LN on US in the nomogram should be further investigated.

Conclusion
In conclusion, the radiomics nomogram had good predictive ability for ALN staging in early stage breast cancer, which could provide incremental information for individual diagnosis and treatment planning based on the current examination.
Abbreviations ALN: axillary lymph node, BMUS: B-mode ultrasound, CI: con dence interval, HER2: human epidermal growth factor receptor 2, IHC: immunohistochemical, MRMR: minimum redundancy maximum relevance, ROC: receiver operating characteristic, SWE: shear-wave elastography, SLN: Sentinel lymph-node Declarations Ethics approval and consent to participate This retrospective study (clinical trial ChiCTR1900027676) was approved by the Institutional Review Board of the participant hospitals, and informed consent was waived.

Consent for publication
Not applicable

Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Competing interests
The authors declare that they have no competing interests.     Data in parentheses are 95% con dence intervals. *P value represents difference of C-index between the radiomics nomogram and clinical model. Figure 1 Flowchart shows inclusion and exclusion criteria.   The calibration curve (training set: a; validation set: b) and decision curve (training set: c; validation set: d) of the nomogram.