In this study, we included 526 patients with DCIS detected by biopsy in the training data set, and the invasive carcinoma underestimation rate was 42.0%, which was consistent with previous reported range [4–10]. Although the underestimation rate is relatively high, the size of invasive cancer is not large, and the vast majority (64.7%) were T1a tumors. This results in only 7.0% patients actually experiencing lymph node metastasis, which was comparable to the 7.2 and 7.3% reported in two studies [6, 24]. This means that 58.0% of patients, whose final pathology was pure DCIS, received overtreatment of lymph nodes, and among patients with upgrading to invasive cancer, only 16.7% (37/221) of lymph node-positive patients truly benefited from lymph node resection. There has been a long-standing controversy on what kind of patient is suitable for axillary lymph node biopsy among those with prior DCIS diagnosis.
Most relevant investigations have explored high-risk factors for underestimation of invasive carcinoma in patients with DCIS diagnosed prior to surgery. Here, we showed that CNB was unable to identify 48.2% (175/363) of invasive carcinomas, compared to 27.5% (46/167) for open biopsy. CNB is an important independent risk factor (p = 0.001 in multivariate analysis) of pathological upgrade. Because of the fact that VABB is not covered by Chinese health insurance, considering the treatment cost, our center did not routinely resort to VABB in preoperative biopsy, which resulted in the lack of this part of data in our study. However, the most widely shared opinion is that VABB is also more accurate than CNB [5, 6, 21, 25]. The reason may be that VABB and open biopsy can obtain more tissue specimens than CNB biopsy. But this does not mean that we encourage more open biopsies. CNB or VABB for suspected malignant lesions should be standardized. We only elaborated on the real situation based on our data, after all, in clinical work, a considerable number of lesions are encountered through open biopsies, such as only calcification on mammography or nipple discharge, patients who have no ultrasound manifestations, and cannot undergo ultrasound-guided core needle biopsy. We also noted that the underestimation proportion for CNB in our study was similar to that in some articles (about 50%) [17, 26, 27] but significantly higher than in other studies in recent years (about 20–30%) [5, 6, 23, 25]. This may be related to the insufficient mammography screening programs in China, as more patients seek medical help due to palpable masses, as high as 80.6% in this study and less than 10% in studies by Ko et al. and Ortega et al. [5, 6]. One study showed only 20% patients had a visible mass on mammography [5], but in the current study, it was up to about 50%. At the same time, several prior studies found that palpability and radiological mass were correlated with upgrade [5, 7–9, 25, 28], as we also observed (both p < 0.05 in univariate analysis). In biopsy pathological features, we demonstrated that moderate -high nuclear grade and elevated Ki-67 levels were significantly related to underestimation of invasive carcinoma, which represent the poor differentiation and high proliferative activity of tumor cells.
We continued to explore high-risk factors for axillary lymph node metastasis in women with DCIS diagnosed by biopsy. Some of the risk factors for lymph node metastasis in our analysis were also reported by previous studies. In the study by Chang [9], the tumor mass according to ultrasound, axillary results on ultrasound, multifocal tumor, surgery, upgrading, and Ki-67 level showed significant correlation with SLN metastasis. In multivariate analysis, Lai et al. [29] found that metastases were associated with large tumor size in imaging before surgery and invasiveness after surgery. In a database analysis [30], a positive lymph node was associated with younger age, increased size, palpability and surgical excisional biopsy. In addition, we also found that mass enhancement on MRI and calcification were significantly associated with lymph node metastasis. It is undeniable that the presence of invasive cancer components was an absolute high-risk factor for positive lymph nodes, and all 45 patients with lymph node metastasis had invasive cancer in our study. The factors according to final pathology, including the expanse of invasive carcinoma and lymphovascular invasion were also associated with positive lymph node [6, 24]. However, these factors cannot be confirmed before surgery, so it is meaningless to determine whether simultaneous lymph node resection is necessary during breast surgery.
Many previous studies have developed multivariate models using routine clinicopathological variables to predict underestimation of DCIS through biopsy diagnosis. The variables included in different models are inconsistent, with AUC ranging from 0.66 to 0.82 [4, 31–33]. Not all studies have conducted external validation of the model. The nomogram developed by Lee et al [32], including DCIS lesion size, hormone receptor expression, nuclear grade, diagnosis on CNB compared with VABB and non-cribriform subtype of DCIS, showed accurately predict invasive disease, the AUC was 0.82 and the validation AUC was 0.70. Study with the largest sample size [4] showed that depending on nuclear grade, mass lesion, multicentric disease and largest linear dimension, the nomogram developed and validated both with c-statistic of 0.71. Our nomogram developed showed AUC was 0.724, AUC was 0.726 after bootstrap, showing good discriminative ability, but external verification was not very satisfactory (AUC = 0.641).
Predicting lymph node metastasis using preoperative characteristics is relatively more difficult, and few studies have specifically evaluated whether DCIS patients have lymph node metastasis and achieve good predictive results (AUC 0.729–0.746) [4, 9, 34]. But in our opinion, including lymph node morphology and final pathology with lymphovascular invasion in the model violates the original intention of creating this model [4, 9]. So far, the model with the largest sample size [34] included age, DCIS detected by screening or not, suspected invasive component at biopsy, palpability and nuclear grade of DCIS and demonstrated good predictive ability. Unfortunately, due to the small number of patients with positive lymph nodes in our study, we were unable to establish an effective predictive model. This is a limitation of our study, and we can only analyze the risk factors for lymph node metastasis as a reference for clinical decision-making.
On the basis of our results and recommendations from existing guidelines, we propose a possible process for axillary management (Fig. 5). For most guidelines that do not recommend routine lymph node biopsy for patients undergoing BCS as a routine practice, selecting patients according to the prediction model with high possibility of upstaging, especially with high lymph node metastasis risk at the same time, may lead to the reduction of a secondary surgery. On the other hand, for guidelines recommending that lymph node biopsy be essential in patients undergoing mastectomy, combining the prediction model and risk factors for positive-lymph node may avoid some excessive lymph node resection. However, this requires great caution and sufficient communication with patients regarding the benefits and risks involved.
Our study did have some restraints. Because of its retrospective nature, selection bias and absent data were possibilities. If more detailed imaging information can be collected, it is believed that our model can be better optimized, especially by utilizing radiomics and machine learning [33, 35, 36]. But currently, this simple model is based on various factors that can be used in daily clinical practice. Because of the different features included in previous studies, we were unable to make comparisons with other models regarding predictive ability. There were relatively few patients in the validation set, that may also be one of the reasons for the unsatisfactory AUC in the validation set.
At last, the external validation set was from the same institute, the nomogram should therefore be tested on other institution independent data sets.