Recently, Associate Professor Tao Tan of Macao Polytechnic University, in collaboration with a research team led by Professor Ritse Mann at the Radboud University Medical Center in the Netherlands, published significant findings in the internationally authoritative journal “npj Breast Cancer” (Impact Factor: 5.9). This study proposes a multi-timepoint breast cancer risk prediction model, representing a breakthrough for personalized breast cancer screening. Associate Professor Tao Tan serves as the corresponding author of the paper, while Xin Wang, a PhD student in Professor Ritse Mann's research group at Radboud University, is the paper's first author.
The research also received contributions from Yuan Gao, Jonas Teuwen, Tianyu Zhang, Anna D'Angelo, Luyi Han, Caroline A. Drukker, Marjanka K. Schmidt, and Regina Beets-Tan from the Netherlands Cancer Institute; Ruisheng Su from Erasmus MC; Jaap Kroes from ScreenPoint Medical; and Nico Karssemeijer from Radboud University Medical Center. This study was supported by the Science and Technology Development Fund of Macao and the Macao Polytechnic University Research Fund, with additional support from the high-performance computing facilities at the Netherlands Cancer Institute.
Breast cancer is the most common malignancy among women worldwide. Traditional age-based mammography screening faces challenges such as high costs, high false-positive rates, and overdiagnosis. Although the concept of "risk-based personalized screening" has been proposed, existing models struggle to balance short-term precise identification with long-term risk assessment. Professor Tan's team, in collaboration with international partners, has innovatively integrated longitudinal imaging information from multiple mammography screenings with clinical risk factors, successfully addressing this industry challenge.
Unlike traditional single-time-point imaging models, the MTP-BCR model can integrate up to five historical screening images, simulating radiologists' diagnostic logic of comparing sequential images to capture subtle risk signals reflecting the temporal evolution of breast tissue. Through multi-level feature integration and a multi-task learning architecture, the model achieves comprehensive risk prediction spanning 1 to 10 years. It demonstrated excellent performance on both the long-term NKI screening cohort from the Netherlands and the external Swedish CSAW-CC dataset.
Notably, the model shows exceptional value in clinical practice: it not only accurately identifies high-risk individuals among healthy populations but also effectively detects potential cancer risks in those traditionally classified as "low-risk," including individuals with benign biopsy results or BI-RADS 0-2 assessments. The model also excels in predicting breast cancer recurrence risk, offering a new tool for postoperative follow-up monitoring. Gradient-weighted visualization analysis confirms the model's ability to consistently localize risk regions, balancing broad risk clue identification with focused lesion attention, demonstrating good clinical interpretability.
This research confirms that spatiotemporal evolution information contained within longitudinal breast images is key to enhancing risk prediction accuracy, providing a novel methodology for precision breast cancer screening. Associate Professor Tao Tan, the corresponding author, stated that the team will further advance prospective multi-center studies to facilitate the model's translation into clinical practice, contributing to a new paradigm of "risk stratification, precise prevention and control" in breast cancer management.