his research group stood out in the 2023 RSNA Breast Cancer Screening Competition and won the silver medal in the global competition, while advancing medical service development through research and technological innovation. RSNA (Radiological Society of North America) is the world's leading conference in medical imaging, representing 31 radiology specialties from 145 countries and regions, leading the global radiology research and application direction. Over 50,000 doctors, researchers, and industry professionals attend RSNA every year. RSNA places great emphasis on the application of artificial intelligence in imaging. To promote the research and application of artificial intelligence in medical imaging, since 2017, RSNA has established a dedicated AI challenge, which has become one of the important competitions in this field. This year, RSNA and Kaggle, the world's largest open data and machine learning competition platform, jointly held a breast molybdenum target detection competition, aiming to develop machine learning algorithms for breast disease screening to improve the accuracy and efficiency of screening.

The global medical imaging competition lasted for three months and attracted nearly 2000 teams and 2146 individuals from dozens of countries around the world. Participants included professionals from well-known medical institutions and top universities at home and abroad (such as the University of Pennsylvania, the University of Tokyo, KAIST, Zhejiang University, Harbin Institute of Technology), as well as medical and artificial intelligence laboratories from institutions such as NVIDIA. Participants were tasked with developing solutions using the comprehensive breast molybdenum target image dataset provided by RSNA in collaboration with multiple imaging laboratories worldwide to quickly and accurately detect patients' breast health based on breast molybdenum target images. This dataset included over 54,000 breast molybdenum target images from more than 13,000 patients. Participants were challenged by noise-labeled information in these datasets, requiring them to effectively handle this difficulty-caused interference while also considering the effectiveness and efficiency of model design within the time limit.

Under the guidance of Professor Tan Tao, the research group conducted in-depth studies of various theoretical and practical methods. They recognized that compared to natural images, breast molybdenum target image detection is more difficult and information aggregation more complex. Many common method designs were difficult to produce positive results in this competition. After multiple discussions and guidance from their professor, they ultimately adopted a method of multi-model, multi-stage, and multi-information fusion. By combining additional semantic information from patients' breast molybdenum target images and fusing the relevance of medical image lesions, they quickly found diagnostic rules for breast molybdenum target images using an effective data sampling algorithm from a large dataset. Their solution ranked in the top 4% on both public and private leaderboards, effectively helping with breast cancer identification, significantly reducing the adverse impact of misdiagnosis on patients, and having the potential to achieve early diagnosis and screening for breast cancer and effectively improve medical procedures and quality.