A study carried out in Sweden and published in The Lancet Oncology has shown that the use of artificial intelligence (AI) in the detection of breast cancer does not increase the risks and may even improve the results. The researchers divided about 80,000 women into two groups, both of whom underwent mammograms, but the first group was evaluated conventionally by two independent radiologists, while the second group was assisted by AI software before a single radiologist analyzed the results. data.
The results revealed that the AI-assisted group did not show worse results compared to the group evaluated by radiologists alone, and even a slightly higher number of cancer cases were detected in the first group. Also, the “false positive” rate, where the initial diagnosis is incorrect, was similar in both groups.
Using AI in this process could cut the workload for radiologists in half, as AI software would only require evaluation by one radiologist. This is especially relevant in countries like France, where breast cancer screening tests are common among women aged 50 to 74.
Although the study results are promising, radiologist Kristina Lang, the study’s lead author, cautions that they are not enough to confirm that AI is ready to be implemented in mammography screening. More research will be needed over the next several years to determine if AI is as effective as human double opinion.
One point to consider is the risk of “overdiagnosis,” that is, the detection of lesions that would not have become dangerous cancers without treatment. Although the study did not provide information on this aspect, other experts note that AI may have overdiagnosed certain forms of early breast cancer.
Despite the challenges and the need for more research, the study has been praised by experts, who stress that reducing the workload of radiologists is of great importance in breast cancer screening programs. The potential for AI to improve this process and assist radiologists in their work is promising, but careful and ongoing evaluation will be required before its widespread implementation.