AI model achieves expert-level accuracy in non-invasive breast cancer diagnosis using MRI data

Researchers from the Hong Kong University of Science and Technology (HKUST) have developed Mixture of Modality Experts (MOME), a large artificial intelligence (AI) model for non-invasive breast cancer diagnosis. Trained on China’s largest multiparametric MRI (mpMRI) breast cancer dataset, MOME achieves expert-level accuracy in classifying tumor malignancy, comparable to that of radiologists with five+ years of experience.

This innovative solution is now undergoing extensive clinical validation across more than ten hospitals and partner institutions, including Shenzhen People’s Hospital, the Guangzhou First Municipal People’s Hospital, and Yunnan Cancer Center, to validate its effectiveness and ensure real-world applicability. The paper is published in the journal Nature Communications.

Harnesses China’s largest mpMRI dataset

Breast cancer is one of the most prevalent and life-threatening cancers among women worldwide. Early detection, accurate molecular subtyping, and the ability to predict patient responses to treatment are crucial in its effective management. While mpMRI provides rich diagnostic information, integrating its diverse imaging modalities (i.e. different MRI sequences) poses challenges for traditional AI systems, especially when sequences are missing in clinical settings.

To address these challenges, the HKUST-led team collaborated with multiple medical institutions to compile the largest Chinese breast mpMRI dataset reported to date and designed MOME, a large AI model capable of learning from diverse types of data.

Using a “mixture-of-experts” framework and a “transformer” architecture, MOME effectively fuses multimodal information and remains robust even when some imaging sequences are missing. The model also supports molecular subtyping and predicts treatment response.

Potential to reduce unnecessary biopsies and enhance treatment predictions

In trial testing, MOME not only demonstrated diagnostic accuracy on par with experienced radiologists but also showed potential in reducing unnecessary biopsies by correctly identifying benign cases among BI-RADS 4 patients—individuals with suspicious breast imaging findings indicating a moderate likelihood of breast cancer (between 2% and 95%).

MOME delivered encouraging results in predicting responses to neoadjuvant chemotherapy, a treatment administered before surgery to shrink tumors and improve surgical outcomes, as well as in subtyping triple-negative breast cancer, a more aggressive subtype that requires specialized treatment strategies.

“MOME’s high adaptability and interpretability hold tremendous potential for integration into clinical workflows. By enhancing diagnostic reliability and decision transparency, MOME highlights the transformative role of AI in medical imaging while enabling for non-invasive and personalized cancer management,” said Prof. Chen Hao, Assistant Professor in the Department of Computer Science and Engineering, the Department of Chemical and Biological Engineering, and the Division of Life Science at HKUST, and one of the corresponding authors of the study.

“With the rapid progress of large AI models and imaging technologies, we believe that models like MOME will play an increasingly vital role in empowering clinicians and improving patient outcomes in the near future,” he added.

The study, titled “A Large Model for Non-Invasive and Personalized Management of Breast Cancer from Multiparametric MRI,” was jointly conducted by HKUST’s Smart Lab, Harvard University, Shenzhen People’s Hospital, PLA General Hospital, and Yunnan Cancer Center. Dr. Luo Luyang, a former postdoctoral fellow of Prof. Chen’s research team at HKUST’s Smart Lab and currently a postdoctoral fellow at Harvard University, served as the first author of the research.