Automated Breast Ultrasound (ABUS)-based radiomics nomogram: an individualized tool for predicting axillary lymph node tumor burden in patients with early breast cancer

Breast cancer is the most common malignant tumor and the main cause of cancer-related death among women [1]. The axillary lymph node (ALN) status determines the need for systemic therapy, the extent of surgery, reconstruction options, and the need for radiation therapy after mastectomy. Accurate assessment of ALN status plays an important role in breast cancer treatment and prognosis.

Based on preoperative assessment, patients without suspicious ALN metastasis by ultrasound (US) should have a sentinel lymph node biopsy (SLNB). Patients with positive ALN metastasis on US are candidates for ultrasound-guided fine needle aspiration (FNA) or core needle biopsy (CNB). For a negative FNA/CNB, SLNB is indicated. If FNA/CNB is positive, axillary lymph node dissection (ALND) is indicated unless neoadjuvant therapy is given. ALND is associated with several complications, including lymphedema (found in up to 25% of women after surgery), infection, shoulder motion restriction, and major vascular and nerve damage [2]. Besides, clinical practice has demonstrated that an important number of patients undergo a secondary ALND when SLNB displays major lymph node involvement [3]. Therefore, accurate preoperative noninvasive assessment of ALN status is essential.

According to the findings of the Z0011 trials, the American Society of Clinical Oncology updated clinical practice guidelines, ALND can be omitted in patients with breast cancer with 1–2 positive SLNs without a decrease in disease-free survival or overall survival [4, 5]. So, the new goal of the ALN status evaluation is tumor burden (low burden, < 3 positive ALNs; high burden, ≥ 3 positive ALNs), instead of metastasis or non-metastasis.

Ultrasound is a widely-used tool in assessing the ALN status preoperatively as it is non-invasive, radiation-free, real-time, rapid, and convenient. Previous studies have proved that axillary ultrasound (AUS) may provide valuable information relevant to ALN status in breast cancer [6]. However, AUS mainly obtains visual image information and focuses on the qualitative analysis of lesions, the diagnostic performances of axillary ultrasound to detect ALN involvement highly depends on the experience of radiologists [3]. Ahmed et al. [7] showed that 43.2% of patients with positive ALN metastasis on US had a low lymph node tumor burden (ALN metastasis < 3). It means that almost half of patients with ALN metastasis positive US assessments are exempt from ALND, which some researchers believe may constitute overtreatment for these patients [8]. AUS may not be a reliable predictor for nodal metastasis [6, 9]. In the age of precision medicine, a more effective and individualized method is urgent to resolve this problem.

As an emerging three-dimensional imaging technique, Automated Breast Ultrasound (ABUS), addresses the limitations of conventional handheld ultrasound (HHUS) and automatically scans the breast based on special high-frequency broadband sensors [10]. Several recent studies [11, 12] have shown that some unique features of ABUS, although they are also visual assessment and qualitative descriptions, may provide additional information for breast lesions. Specifically, retraction phenomenon appears as a satellite model around the lesion, with high sensitivity (80%–89%) and specificity (96%–100%) for breast cancer [13‐15]. Radiomics extracts high-throughput quantitative features that may not be directly observable with the naked eye from single or multiple medical images. Radiomics has been more recently applied to distinguish benign malignant breast lesions [16], predict lymph node status [17, 18], and even evaluate treatment response [19]. According to the radiomics quality score proposed by Lambin et al. [20], ABUS images with standardized, repeatable, and high-resolution characteristics would be fit for radiomics analysis. However, to our knowledge, there is no ABUS-based radiomics study to differentiate ALN tumor burden. For patients with high tumor burden can assist in identifying what kind of initial axillary surgery can overlook the SLNB and undergo ALND specifically and assist in employing neoadjuvant chemotherapy or individualized adjuvant radiotherapy. For patients with low ALN tumor burden, unnecessary treatment, and potential complications due to surgery can be avoided.

In the study, we constructed radiomics score based on ABUS image for predicting ALN tumor burden, which AUC values were 0.794 in the training set and 0.789 in the test set. The ABUS model, integrating diameter, hyperechoic halo, and retraction phenomenon, showed better performance, with AUC values of 0.772 in the training set and 0.736 in the test set. By multivariate logistic regression analysis, we developed a radiomics nomogram integrating the US-reported ALN status, retraction phenomenon and radiomics score, which showed best performance with AUC values 0.876 in the training set and 0.851 in the test set.

ABUS is a breakthrough in breast ultrasound that has been developed to meet the limitations of HHUS. ABUS provides three-dimensional imaging of the entire breast with multiplanar reconstructions, which has been demonstrated to improve diagnostic accuracy [24]. Several studies [11, 25‐31] have focused on the ability of diagnosing breast lesions and assessment of response to neoadjuvant chemotherapy by ABUS. These studies indicated good diagnostic performance of ABUS. Recently, several studies [12, 32] showed some ABUS imaging features were significantly correlated with ALN status. In our study, ABUS model displayed adequate discriminative ability (AUC 0.772 in the training set and 0.736 in the test set). Three ABUS imaging features, diameter, hyperechoic halo, and retraction phenomenon, were obviously associated with ALN tumor burden (all P < 0.05), in this study. Some studies [33] reported that the size of primary breast cancer was significantly related to ALN status. Larger breast cancers have a wider range of glands invaded by cancer cells, and the probability of ALN metastasis through lymphatic drainage is also higher. In our study, retraction phenomenon was identified to be the strongest independent predictor with high diagnostic performance in differentiating high and low ALN tumor burden(P < 0.001). Desmoplastic reaction of breast malignancy, which can produce contraction of the surrounding tissues toward the mass and disrupt normal parallel tissue planes, might help explain the generation of retraction phenomenon [34]. The appearance of the retraction phenomenon indicates that breast cancer is highly aggressive and has poor response to treatment [35]. Jiang et al. [36] and Tang et al. [37]showed that the smaller and more superficial invasive carcinomas with lower histological grades tended to present with retraction phenomenon. A hyperechoic halo, also known as converging pattern, is caused by the compressed fibrous surrounding tissue or the infiltration between the tumor and the surrounding tissue [38, 39]. It reflects the degree of invasion of cancer cells and may be an important indicator of poor prognosis. Similarly, Tang et al. [37] reported that the malignant masses were associated with retraction phenomenon and discontinuous hyper-and hypoechoic rim (p < 0.001 for each).

However, visual assessment and qualitative descriptions for US features extremely rely on the personal clinical experience and subjective judgment of radiologists. Gillies et al. [40] have illustrated that tumor characteristics at the genetic and cellular levels can be captured from medical images by extracting and computing high-throughput features. Qiu et al. [41] extracted radiomics features from B-mode ultrasound and integrated the significant clinical characteristics of patients to construct radiomics model for predicting ALN metastasis. Therefore, this study comprised the radiomics signature and US-reported ALN status, with AUC values of 0.816 in the training cohort and 0.759 in the validation cohort. Similarly, Gao et al. [42] established a nomogram based on radiomics analysis of primary cancer B-mode ultrasound for predicting ALN tumor burden, with AUC values of 0.846 in the training cohort and 0.733 in the validation cohort. The results of above research were similar to our results. Jiang et al. [43] extracted radiomics features from shear-wave elastography and B-mode ultrasound and integrated Clinical characteristics of patients to construct radiomics model. The research showed in the performance of discriminating disease-free axillary (N0) and any axillary metastasis (N + (≥ 1)), it achieved a C-index of 0.845 for the training cohort and 0.817 for the validation cohort. The tool could also discriminate between low (N + (1–2)) and heavy metastatic ALN burden (N + (≥ 3)), with a C-index of 0.827 in the training cohort and 0.810 in the validation cohort.

ABUS provides standardized scanning protocols and uncouples detection from image acquisition, hence improving reproducibility, reducing operator dependency and radiologist workload [10]. To our knowledge, there is no published study that has showed whether the value of radiomics score from ABUS would benefit prediction of ALN tumor burden in early breast cancer. In our study, 13 radiomics features from axial images were screened out and used to established radiomics score. The ABUS radiomics score showed moderate discriminative ability (AUC 0.794 in the training set, 0.789 in the test set).

We constructed a radiomics nomogram that integrated radiomics score with retraction phenomenon and US-reported ALN status to improve its predictive accuracy for ALN tumor burden. Based on the proposed risk classifier, the ABUS radiomics nomogram was able to classify patients into low- and high-tumor burden groups. According to ACOSOG Z0011 trials, we should pay more attention to the ALN tumor burden, instead of metastasis or non-metastasis. The radiomics nomogram resulted in 17.4% (30/172) and 17.5% (14/80) false positives in the training and test sets, which was obviously below the rates in US-reported ALN status (51.1% (88/172), 43.8% (35/80) in the training and test sets). The results showed that the radiomics nomogram could evidently reduce the false positive rate, compared with US-reported ALN status. Accurate forecasting of ALN high tumor burden can assist in identifying what kind of initial axillary surgery can overlook the SLNB and undergo ALND specifically and assist in employing neoadjuvant chemotherapy or individualized adjuvant radiotherapy [44]. Besides, for patients with low ALN tumor burden, unnecessary treatment, and potential complications due to surgery can be avoided. Figure 7 showed an example, the prediction of ALN tumor burden in patient.

Zhou et al. [45] firstly used deep learning methods in the evaluation of ALN status. Which included 974 primary tumor images of 756 patients with lymph node-negative early breast cancer, and validated the model in an independent validation set of 78 patients, with an AUC of 0.89. However, the logic behind the features and decisions based on deep learning is still a “black box”, which means the results based on deep learning are difficult to interpret [46]. On the contrary, in our study, “wavelet.LHH_glcm_JointEnergy”, “wavelet.HLL_firstorder_Kurtosis”, “wavelet.HHL_firstorder_Skewness”, and “wavelet.HHL_glcm_SumEntropy” were obviously associated with ALN tumor burden, indicating that these features strongly reflect ALN tumor burden. In our study, most of selected radiomics features are texture features after wavelet transform. Wavelet transform, which measures the resolution of image signals in different time, space, and frequency scale planes, is extremely useful for replaying even subtle but important texture information that is ignored by radiologists in low-contrast US images. Previous researches [19, 47] have demonstrated the texture features after wavelet transform are used to construct a prediction model.

There are several limitations to our study. First, the study was a retrospective study, inevitably there existed sample bias. Second, the sample originated from a single center and lacked external validation. Therefore, the number of positive patients was 58 (58/310,41/213 in the training set and 17/97 in the test set). The difference between the two sets may be caused by the small number of positive patients. In the future, more patients (particularly more positive patients), multi-center samples, and external validation are needed. Third, the three-dimensional ROI was manually delineated on the axis ABUS image by the radiologist. However, manually delineating the ROI was extremely time-consuming and inevitably involved inter-observer variability. In the future, we should research semi-automatic or automatic segmentation algorithms to overcome the above problems. Finally, we extracted radiomics features from intratumoral regions and failed to exploit peritumoral radiomics features in the study. The radiomics features of intratumoral and peritumoral regions should be integrated into further research.

The ABUS radiomics nomogram showed favorable ability for predicting ALN tumor burden, which may provide additional benefits in treatment strategies for patients with early breast cancer, especially for patients with low tumor burden. With a more individualized and precise assessment for ALN tumor burden, radiomics nomogram will assist clinicians to make the optimal treatment strategy and avoid overtreatment.