This retrospective study was based on the existing medical records of patients with luminal-like breast cancer in the breast cancer registry and raw mammographic data at the Changhua Christian Hospital from 2011 to 2015. A total of 1941 ER-positive patients were screened. The participants were aged 35-75 years and having full record of mammographic exam during whole adjuvant HT treatment. The patient’s age, calculated BMI (in kg/m2), tumor size, grade, stage, lymph node status, estrogen receptor (ER)/progesterone receptor (PR)/human epidermal growth factor receptor 2 (HER2) status, and MD grade before HT were obtained for all participants from their medical records. The exclusion criteria include (a) non-luminal-like patient (ER negative), (b) not administrated with HT after confirmed diagnosis, (c) had bilateral breast or occult breast cancer or any other condition which was not appropriate for evaluation, and (d) underwent treatment for less than 48 months. To prevent over express the treatment effect of tamoxifen on MDR, patients with T0/Tis/T1 stage disease in TNM classification were also excluded. Treatment with adjuvant HT was generally decided according to the institution’s guidelines. The standard duration of treatment with tamoxifen is 5 years. Postmenopausal women were treated with the aromatase inhibitors anastrozole and letrozole for up to 5 years after surgery or after 2 to 3 years of tamoxifen. A detailed flow of patient selection is shown in Figure 1.
In this study, the MD measurement was utilized the full-field digital mammography (FFDM) combining automated mammographic density estimated software. It minimizes the bias, compared to traditional human visual assessment. All participants were undergoing the same procedure of institute and had four-view digital mammograms in each exam. Standard two-dimensional digital mammograms were obtained for all the participants using mainstream FFDM systems, including Senographe Essential/Senographe DS (GE Medical Systems, Milwaukee, WI, USA), Mammomat Inspiration (Siemens AG Healthcare, Erlangen, Germany), and Selenia Dimensions systems (Hologic Inc, Bedford, MA, USA). The MD measurements, including the mammographic density grade (VDG), overall dense breast volume (in cm3), and volumetric density percentage (VPD), were estimated and quantitated using the VolparaTM software version 1.5.1 (Volpara Health Technologies, Wellington, New Zealand). To decrease the effect of tumor tissues on the density estimates and to obtain more accurate results, the VPD in the present study were calculated only from the contralateral lesions. The change in the MD at the last measurement compared to that at baseline was used to calculate the MDR. It is note that the VDG in VolparaTM have different density grade estimation method to the Breast Imaging Reporting and Data System (BI-RADS). However, it is proved the performance and accuracy of density grade estimation was equal to experienced radiologist. We do not repeat the mammographic breast density measurement of consistent between radiology and VDG again.
Mammographic density reduction (MDR) was based on two digital mammograms; the first was the baseline mammogram which taken before surgery and adjuvant HT (preMD), and the second was 6 to 18 months after the start of adjuvant HT (postMD). It defined as the absolute MDR of these two images:
MDR(%) = preMD - postMD/100.
The diagnostic accuracy of MDR threshold was estimated using the area under the receiver operating characteristics (ROC) curve (AUC) and was compared with DeLong’s nonparametric test. The Youden’s index  was utilized to determine the optimal cut-off point for the specificity and sensitivity result.
Cox proportional hazards model was used to estimate the disease-free survival (DFS) rate and survival curve. We also identified the different factors that are associated with MDR on DFS and hazard ratios (HRs) to represent the risks through Cox proportional hazards model. The overall mode fit estimation in Cox regression model was utilized the Chi-squared statistic tests the relationship between time and all the covariates in the model, to ensure the proportional hazard assumption is satisfied. The p-value less than 0.05 was considered significant in statistic and in model fit. Thee statistical analyses were performed using MedCalc for Windows (Version 19.5.3, MedCalc Software, Ostend, Belgium).