AI rivals radiologists in detecting breast cancer

AI systems can detect breast cancer just as well as radiologists, according to a study published March 5 in the Journal of the National Cancer Institute.

In an analysis of nine multi-reader, multi-case study datasets comprising more than 2,500 exams, Ioannis Sechopoulos and co-authors found robotic systems were noninferior to 101 radiologists in interpreting breast scans and identifying possible lesions. According to the team, computer-aided detection systems have been helping detect and classify breast lesions for around three decades now, but no studies have proven those systems are more cost-effective or accurate than radiologists themselves.

“AI systems performing at radiologist-like levels in the evaluation of digital mammography would improve breast cancer screening accuracy and efficiency,” Sechopoulos et al. wrote in their report. “We aimed to compare the stand-alone performance of an AI system to that of radiologists in detecting breast cancer.”

Each dataset the authors considered consisted of digital mammography exams acquired with systems from four different vendors, assessments from multiple radiologists and ground truth verified by histopathological analysis or follow-up. The project yielded a total of 2,652 exams—652 malignant—and 28,296 independent interpretations by 101 radiologists.

Sechopoulos and colleagues measured the radiologists’ results against those of a commercially available AI system, which analyzed each exam to determine its level of suspicion of cancer present on a scale from 1 to 10. Detection performance between the radiologists and AI system was compared using a noninferiority null hypothesis at a margin of 0.05.

The authors found the performance of the AI system was statistically noninferior to that of the average of all 101 radiologists—the AI system had a 0.840 area under the ROC (receiving operating characteristic) curve and radiologists scored a 0.814. In radiology, the ROC is used to represent the effects of varying levels of sensitivity and specificity.

Sechopoulos et al. also reported that in 61.4 percent of cases, the AI system had an area under the curve higher than that of radiologists.

“Although promising,” the authors wrote their results indicate the need for further investigation to test the feasibility of using AI for cancer detection in real-world settings.

“Before we could decide what is the best way for AI systems to be introduced in the realm of breast cancer screening with mammography, we wanted to know how good can these systems really be,” Sechopoulos said in a press release. “It was exciting to see that these systems have reached the level of matching the performance of not just radiologists, but of radiologists who spend at least a substantial portion of their time reading screening mammograms.”