Very efficient algorithms from the last RSNA AI…

The RSNA Screening Mammography Breast Cancer Detection AI Challenge was a participative competition organized at the 2023 RSNA congress, with more than 1,500 participating teams. In addition, It is the subject of a Article in Radiology reviewwhich analyzes the performance of algorithms involved in this context.

A challenge that has generated highly efficient algorithms for breast cancer screening

« We were impressed by the number of participants. Moreover, by the number of AI algorithms submitted as part of the challenge, explains Professor Yan Chen, cancer screening professor at the University of Nottingham (United Kingdom). Consequently, It is one of the most popular IA challenges of the RSNA. Therefore, We have also been impressed by the performance of algorithms. In addition, taking into account the relatively short time allocated for their development and the very efficient algorithms last rsna need to obtain open source learning data. Nevertheless, »

The objective of the challenge was to find AI models to automate cancer detection during screening mammograms. help radiologists work more effectively, improve the quality and safety of patient care, and potentially reduce unnecessary costs and medical procedures.

Tools tested on more than 10,000 mammograms by a team of researchers

The Emory University of Atlanta (Georgia – USA) and BreastScreen Victoria (Australia) provided for this challenge a set of training data of around 11,000 breast cancer screening images.

Professor’s research team evaluated 1. 537 operational algorithms subject to the challenge, testing them on a set of 10,830 different mammograms of the entry data and whose results of anatomopathology confirmed the presence or the non-presence of cancer. In total, the algorithms obtained median specificity rates of 98.7% to confirm the very efficient algorithms last rsna absence of cancer on mammograms, sensitivity of 27.6% to positively identify cancer and recall rate of 1.7%. By combining the 3 and 10 most efficient algorithms, the researchers obtained a sensitivity of 60.7 % and 67.8 % respectively.

The performance of the 10 best algorithms close to those of an average radiologist

« When assembling the most efficient algorithms. we were surprised to see the complementarity of the different AI algorithms, making it possible to identify different cancerscontinues Professor Chen. The algorithms had optimized thresholds for a positive predictive value. high specificity, so that different cancerous characteristics on different images triggered high scores differently depending on the algorithms. »

According to the researchers. the creation of a set of the 10 most efficient algorithms produced performance close to those of an average screening radiologist in Europe or Australia. Overall, algorithms have shown greater sensitivity for the detection of invasive cancers very efficient algorithms last rsna than for non -invasive cancers.

Additional studies in the project for inclusion in clinical practice

Like many Participants AI models are open source, the results of the challenge could contribute to improving experimental and commercial AI tools for mammography. “” By making algorithms. a full set of imaging data, participants provide precious resources that can supply research and allow the comparative analysis necessary for the effective and safe integration in clinical practice in clinical practice She adds.

The research team plans to conduct additional studies to compare the performance of the main challenge algorithms to those of the products available on the market. using a larger and more diverse set of data. It also plans to assess the effectiveness of smaller. more difficult test sets with robust human reading benchmarks to guarantee the quality of radiologists’ performance as an AI very efficient algorithms last rsna assessment.