The institutional review board (IRB) of Severance Hospital, Yonsei University (IRB approval No: 4-2022-0519) approved this retrospective study and waived the requirement for informed consent based on its study design.

We searched our institutional database for women who were diagnosed with DCIS via percutaneous biopsy, including core needle biopsy and imaging-guided vacuum-assisted biopsy (VAB). From January 2015 to December 2019, 743 DCIS were diagnosed in 717 women via image-guided percutaneous biopsy. The exclusion criteria were as follows: (1) women who were surgically treated for breast cancer in the ipsilateral breast (n = 58), (2) women who had invasive cancers diagnosed in the ipsilateral breast (n = 78), (3) women who underwent neoadjuvant chemotherapy due to invasive cancers in the contralateral breast (n = 33), (4) women who were lost to follow-up after DCIS diagnosis (n = 61), and (5) women who only had analog mammograms from outside hospitals that were inadequate for AI-CAD analysis (n = 73) (Fig. 1).

Flowchart of patient selection

The clinical characteristics of patients including age, family history of breast cancer, personal history of breast cancer, presence of bilateral breast cancer, presence of related symptoms such as palpability or nipple discharge, method of percutaneous biopsy, and tumor size were extracted from our electronic medical record (EMR) system.

One of two dedicated digital mammography units was used for the mammography examinations (Senographe DS, GE Medical Systems; Lorad Selenia, Hologic). Standard mediolateral oblique (MLO) and craniocaudal (CC) mammograms and magnification views with 90° lateral and craniocaudal projections, if required, were obtained for all patients.

One board-certified, breast-dedicated radiologist with 14 years of experience in breast imaging (J.H.Y.) retrospectively reviewed the baseline mammograms that were collected routinely before biopsy. Mammographic features of abnormalities that correlated to the biopsy-proven DCIS were categorized into the following four categories: (1) mammographically occult (DCIS detected on supplemental ultrasound (US)), (2) mass/asymmetry/distortion, (3) calcifications only, and (4) combined mass/asymmetry/distortion with calcifications (referred to as “combined features”). Final assessments according to the American College of Radiology Breast Imaging Reporting And Data System (ACR BI-RADS) [21] were also determined by the radiologist during the retrospective review. The radiologist was blinded to the final surgical diagnosis.

A commercially available AI-CAD algorithm (Lunit INSIGHT for Mammography, version 1.1.4.3, Lunit Inc., Seoul, Korea) that was previously validated through a multinational study [22] was used for analyzing mammograms. The algorithm, based on the ResNet-34, a popular deep convolutional neural network (CNN) architecture, was trained using 31,604 cancer-positive mammograms and 19,625 benign mammograms with pixel-label labels indicating lesion locations annotated by 12 breast-dedicated radiologists. The algorithm provides region of interest (ROI) marks for abnormalities on mammograms while providing corresponding abnormality scores (referred to as AI-CAD scores, ranging 0–100%) per view.

In this study, we employed a three-pronged approach for AI-CAD scores: (1) numerical AI-CAD score provided in raw numbers (ranging from 0 to 100%); (2) AI-CAD scores dichotomized into < 50% and ≥ 50%; and (3) graded AI-CAD score of < 25%, 25–50%, 50–75%, and ≥ 75%.

Information regarding nuclear grade (low, intermediate, or high grade) and presence of comedonecrosis was collected from the pathology reports from percutaneous biopsy. Tumors on percutaneous biopsy specimens were histologically classified using the World Health Organization criteria [23].

Ground truth in terms of pure DCIS or invasive cancer was confirmed after surgery. The Shapiro-Wilk test and Kolmogorov-Smirnov test were performed to test for normality for age, tumor size, and AI-CAD scores. As the normality assumption was not satisfied, the median values for these factors were calculated and compared. The Mann-Whitney U test was used to compare clinicopathological variables between pure DCIS and DCIS with invasive upgrade. Mammographic variables such as imaging features on mammography, ACR-BI-RADS final assessment, and median AI-CAD score were also compared between pure DCIS and DCIS with invasive upgrade using the Mann-Whitney U test and Fisher’s exact test.

Univariable logistic regression analysis using clinicopathological variables and mammographic variables was performed to assess predicting factors for invasive upgrade in DCIS. Subsequently, multivariable logistic regression analysis was performed to identify independent predictive mammographic variables after adjusting for clinicopathological variables. Variables with p values less than 0.05 in the univariable logistic regression analysis were included for multivariable logistic regression analysis. The predictability of the multivariable models was evaluated with the area under the receiver operating characteristics curve (AUROC). A subgroup analysis was conducted specifically on DCIS detected on mammography, referred to as “mammographically detected DCIS.” This analysis excluded cases that were mammographically occult, to simulate situations where supplemental screening with imaging modalities other than mammography is not common.

All statistical analyses were performed using SAS (version 9.4, SAS Inc.). p-values less than 0.05 were considered statistically significant.