Deep Learning-Based Risk Stratification of Ductal Carcinoma In Situ Using Mammography and Abbreviated Breast Magnetic Resonance Imaging

Tingfeng Zhang^1,2,3Tingting Cui⁴Zhenjie Cao^5,6Jintao Hu^7,8Jie Ma^9*

• ¹First Affiliated Hospital of Southern University of Science and Technology, Shenzhen, China
• ²Division of Breast Surgery, Department of General Surgery, Shenzhen People’s Hospital, shenzhen, China
• ³The Second Clinical Medical College, Jinan University, Shen zhen, China
• ⁴Key Laboratory of Biomedical Engineering of Guangdong Province, South China University of Technology, guangzhou, China
• ⁵Shenzhen International Graduate School, Tsinghua University, Shenzhen, China
• ⁶PingAn Tech, Shenzhen, Shenzhen, China
• ⁷Department of Pathology, Shenzhen People's Hospital, Jinan University, Shenzhen, China
• ⁸Department of Anatomical and Cellular Pathology, Prince of Wales Hospital, The Chinese University of Hong Kong, Hong Kong, China
• ⁹Department of Radiology, Shenzhen People’s Hospital, The Second Clinical Medical College of Jinan University, Shenzhen, China

BackgroundCurrent management of ductal carcinoma in situ lacks robust risk stratification tools, leading to universal surgical and radiotherapy interventions despite heterogeneous progression risks. Optimizing therapeutic balance remains a critical unmet clinical need.Materials and MethodsWe retrospectively analyzed two patient cohorts. The first included 173 cases with BI-RADS category 3 or higher findings, used to compare the diagnostic accuracy of four abbreviated MRI protocols against the full diagnostic MRI. The second cohort involved 210 patients who had both mammography and abbreviated MRI. We developed two separate predictive models—one for pure ductal carcinoma in situ and another for invasive ductal carcinoma with associated ductal carcinoma in situ—by integrating clinical, imaging, and pathological features. Deep learning and natural language processing techniques were used to extract relevant features, and model performance was assessed using bootstrap validation.ResultsAbbreviated Magnetic Resonance Imaging protocols demonstrated similar diagnostic accuracy to the full protocol (P > 0.05), offering a faster yet effective imaging option. The pure group incorporated features like nuclear grade, calcification morphology, and lesion size, achieving an Area Under the Curve of 0.905, with 86.8% accuracy and an F1 score of 0.853. The model for invasive cases incorporated features Ki-67 status, lymph vascular invasion, and enhancement patterns, achieved an Area Under the Curve of 0.880, with 86.2% accuracy and an F1 score of 0.834. Both models showed good calibration and clinical utility, as confirmed by bootstrap resampling and decision curve analysis.ConclusionDeep Learning-driven multimodal models enable precise ductal carcinoma in situ risk stratification, addressing overtreatment challenges. abbreviated Magnetic Resonance Imaging achieves diagnostic parity with full diagnostic protocol, positioning Magnetic Resonance Imaging as a viable ductal carcinoma in situ screening modality.