The accuracy of digital mammograms is no different, regardless of whether a human or a commercially available AI-based system reads them.

That’s the claim made in a new study that found that AI-based readings are equivalent to those of human radiologists, suggesting that greater implementation of such systems is necessary and could prove helpful in reducing large numbers of patients waiting to be screened.

"In some countries, there is an alarmingly high number of breast screening radiologists set to retire in the next few years. For example, in the UK, one in five breast radiologists will retire by 2020 and one in three by 2025. In general, the incoming radiologists are not in numbers that are expected to replace the retiring ones," Ioannis Sechopoulos, an associate professor of radiology and nuclear medicine at AXTI laboratory and one of the paper's authors, told HCB News. "We are interested on how we can optimize breast cancer screening with mammography and breast tomosynthesis, the newest breast screening technology."

Breast cancer is responsible for approximately 500,000 deaths worldwide annually, an issue that has prompted clinicians to investigate other potential screening methods. While in development since the 1990s for breast lesion detection in mammography, no studies have shown improvements in performance or cost effectiveness with AI-based systems, hindering their adoption as a method, according to the study authors.

The hope of many is that use of these systems could potentially reduce the high labor intensity that current screening programs face due to the large number of women who require examinations.

The study examined the use of AI at a case level, comparing the performance of the Transpara system by ScreenPoint to that of 101 radiologists who scored nine distinct cohorts of mammograms from four different manufacturers as part of studies performed in the past for other reasons. A total of 2,652 exams were assessed and broken down into data sets, each of which consisted of findings that came from systems of four different vendors, and were assessed by multiple radiologists per exam. The combined number of interpretations by the 101 radiologists was 28,296.

The findings of the system were found to be statistically non-inferior to that of the average radiologist out of the 101, achieving a cancer detection accuracy on par with an average breast radiologist in the retrospective setting. Sechopoulos says that while the results are promising, additional steps are required and more variables must be assessed.

"We need to evaluate the performance of the systems with datasets that are not cancer-enriched, like they were in this study," he said. "In other words, in true screening datasets. This is possible and we are planning on doing this, but it is a large effort due to the low number of cancer cases present in a real set of screening cases. So spanning such a large variety of sources and types of images and radiologist readings like we did for this study is very challenging. We also want to determine what is the best way to incorporate AI into the screening process. There are many variables to consider and evaluate in the aim of optimizing both the screening performance and its cost-effectiveness."

He adds that greater trust in AI is also essential "For this approach to be incorporated in one way or another, we need everybody involved in the screening process, including, and most importantly, the women undergoing screening, to accept the use of AI in the evaluation of the mammograms, especially if it will be used as the sole reader for some percentage of cases. This may sound scary to some people now, but we're showing that it should not be a scary thought. Medico-legal aspects need to be settled, such as when the AI system misses one cancer, who is responsible. These are the same issues being faced in all other aspects of life were AI is being introduced, like driver-less cars."

The findings were published in Journal of the National Cancer Institute.