AI can enhance accuracy of breast cancer screening, study finds

In Clinical Studies News by Barbara Jacoby

By: Brian P. Dunleavy
Artificial intelligence-assisted mammogram screening is more accurate at identifying women with breast cancer than assessment by a radiologist, a study published Thursday by JAMA Oncology found.

However, the most effective diagnostic approach combined computer algorithms with an evaluation by a radiologist, which correctly identified women with the disease 85% of the time, the researchers said.

“There is now evidence that AI has reached a performance on par with radiologists — as an independent reader of screening mammograms,” study co-author Dr. Fredrik Strand told UPI.

“I am convinced that AI will be part of the future of breast cancer screening,” said Strand, who is a researcher in oncology and pathology at the Karolinska Institute in Sweden.

AI is the design of so-called “smart machines” that are capable of performing tasks that require human intelligence.

In the case of breast cancer screening, AI computer-aided detection algorithms might be useful in reducing the workload for radiologists and limiting “the broad variation in performance among human readers,” Strand and his colleagues said.

Current computer-aided detection systems can perform two separate roles: as an assistant directing the radiologist’s attention to suspicious areas in the mammogram and as an independent reader making an assessment of the whole examination without radiologist intervention, the researchers said.

For this study, the Karolinska Institute researchers assessed three AI-based computer-aided detection systems by using data from 739 Swedish women screened for breast cancer between 2008 and 2015 and from 8,066 healthy controls, and compared the findings to those from 45 experienced radiologists.

Twenty-five of the radiologists served as “first-readers” — or initial evaluators of mammogram findings — while the remainder were “second readers,” or essentially a second set of eyes on the findings, the researchers said.

Without input from a radiologist, the best-performing algorithm was able to correctly identify women with breast cancer 81% of the time, while the other two systems did so 67% of the time, the data showed.

In comparison, first-reader radiologists correctly identified women with the disease 77% of the time, while second readers did so 80% of the time, according to the researchers.

In addition, all three algorithms were able to correctly identify women who did not have disease roughly 97% of the time, or about the same as first- and second-reader radiologists, the researchers said.

However, the most efficient approach was combining a computer-aided detection algorithm with a human radiologist. That approach correctly identified women with the disease 85% of the time and ruled out women without the disease 99% of the time, the data showed.

“To reach the highest level of cancer detection, one radiologist should also read the mammograms,” Strand said. “Radiologists are still needed to maximize cancer detection and to discuss suspicious cases.”