Rsna AI Challenge models show excellent performance to detect breast cancer in mammograms

The algorithms submitted for an AI challenge hosted by the Radiological Society of North America (RSNA) have shown excellent performance for the detection of breast cancers in mammography images, increasing sorting sensitivity while maintaining low recall rates. RadiologyThe RSNA Premier Journal.

The detection of RSNA AI breast breast cancer was a competition in 2023, with more than 1,500 groups participating. THE Radiology The article describes an analysis of the performance of algorithms, led by Yan Chen, Ph.D., a professor in cancer control at the University of Nottingham in the United Kingdom.

We were overwhelmed by the volume of contestants and the number of AI algorithms submitted as part of the challenge. It is one of the most participants-in RSNA AI challenges. We were also impressed by the performance of the algorithms that received the relatively short window allowed for the development of the algorithm and the requirement to derive training data from open origin. “

Yan Chen, Ph.D., Professor in Cancer View, University of Nottingham

The purpose of the challenge was to supply AI models that improve the automation of cancer detection in sorting mammals, helping radiologists work more effectively, improving the quality and safety of patient care and possibly reducing costs and unnecessary medical procedures.

RSNA called for participation from teams around the world. Emory University in Atlanta, Georgia and Victoria of Victoria in Australia provided a set of about 11,000 breast visibility images and challenge participants could also supply public training data for their algorithms.

Professor Chen’s research team evaluated 1,537 work algorithms that have been challenged, testing them in a total of 10,830 Single Spile Exams-separately from the training set-confirmed by the results of the pathology as positive or negative for cancer.

Overall, algorithms yielded 98.7% specialization for the confirmation of cancer that did not exist in mammography images, 27.6% sensitivity to positive cancer determination and a percentage of recall-percentage of cases that consider 1.7% positive. When the researchers combine the 3 and the top algorithms they perform, it reinforced the sensitivity to 60.7% and 67.8% respectively.

“When we set up the top entries, it was surprised that different AI algorithms were so complementary, identifying different cancers,” Professor Chen said. “The algorithms had thresholds that were optimized for positive prognostic value and high specialization, so different characteristics of cancer in different images caused high ratings differently for different algorithms.”

According to the researchers, the creation of a total of 10 best performance algorithms produced a performance close to that of an average radiologist in Europe or Australia.

Individual algorithms have shown significant differences in performance depending on factors such as the type of cancer, the manufacturer of the imaging equipment and the clinical space where the images were acquired. Overall, algorithms were more sensitive to detecting invasive cancers than for non -invasive cancers.

Since many AI models of participants are open sources, the results of the challenge can help further improve both the experimental and AI commercial tools for mammography, with the aim of improving the effects of breast cancer worldwide, Professor Chen explained.

“With the liberalization of algorithms and a comprehensive set of public imaging data, participants provide valuable resources that can lead to further research and allow the comparative evaluation required for the effective and safe integration of AI into clinical practice,” he said.

The research team plans to conduct surveillance studies to compare the performance of the Top Challenge algorithms against commercially available products using a larger and more different data set.

“In addition, we will explore the effectiveness of smaller, more challenging test sets with powerful human reader reference indicators-such as those developed by the Performs program, a UK-based program for evaluating and safeguarding the quality of the Radiologists as a Radiologists as an AI approximation. Prof. Chen.

RSNA hosts an AI challenge annually, with this year’s contest looking for submissions for models that help detect and detect intracranial aneurysms.