Background: The preoperative diagnosis of ductal carcinoma in situ (DCIS) by core needle biopsy (CNB) can be upstaged in the final pathology, and this possibility is linked to the controversy over whether axillary staging is necessary in primary operation. In this study, we developed an artificial intelligence (AI)-powered Imaging Biomarker in Mammography (IBM) that can predict tumor invasiveness in preoperative mammography and evaluated its performance in an external validation cohort. Methods: A total of 151,764 exams of 4-view mammograms were collected from five institutions of three countries to develop the AI algorithm for breast cancer detection, where 31,776 were cancer exams. In previous studies, the performance of this breast cancer detection algorithm has already been evaluated, and in this study, we further developed the AI-powered IBM for predicting tumor invasiveness on top of the AI algorithm for breast cancer detection. To develop the AI-powered IBM for predicting invasiveness, final diagnosis information was collected for 8,251 cancer exams (472 DCIS, 388 ductal carcinoma in situ (DCIS-MI), and 7,391 invasive ductal carcinoma (IDC)), and 886 cancer exams (44 DCIS, 51 DCIS-MI, 791 IDC) were additionally collected for internal validation. The AI-powered IBM was developed via two stages of training – 1) training with diagnosis labels (cancer vs non-cancer), followed by 2) fine-tuning with invasiveness labels (DCIS, DCIS-MI, IDC). The AI-powered IBM also tested in an external validation cohort of 699 cancer exams (68 DCIS, 19 DCIS-MI, 612 IDC) and all the exams were confirmed by surgical biopsy. Results: The AI-powered IBM showed an area under the curve (AUC) values of 0.968 for breast cancer detection and it successfully distinguished IDC from DCIS and DCIS-MI with AUC values of 0.898 and 0.851, respectively (Table). In addition, the AUC value in terms of discriminating between DCIS and DCIS-MI was 0.752. When the AI-powered IBM was tested in the external cohort, it could detect breast cancer with the AUC of 0.952 and, its performance in terms of invasiveness prediction was similar that of the internal validation (IDC vs DCIS, 0.810; IDC vs DCIS-MI, 0.846), which supports the AI-powered IBM is applicable to the unseen mammography exam. Conclusions: The AI-powered IBM can distinguish IDC from DCIS and DCIS-MI in mammography. The results support that the AI-powered IBM can be used as a biomarker to help determine the surgical plan that includes whether or not to perform the axillary dissection.