AI as a second reader outperforms radiologists
Artificial intelligence may be more well equipped to perform as a second reader on breast cancer screenings than radiologists.
New research published in Lancet Digital Health details the improved cancer detection rates achieved when AI serves as a second reader. According to the study, use of the tool improves sensitivity, especially in cases involving interval cancers.
“On average, interval cancers are more lethal and are associated with more aggressive biological factors,” Suzanne L van Winkel, MSc, with the medical imaging department at Radboud University Nijmegen Medical Center, Netherlands, and colleagues noted. “20–30% of the malignant lesions that are detected in the screened population are in retrospect already visible on previous mammograms, despite not being recalled at that time. Thus, not all cancers that are present in the screened population are recognized during a screening examination.”
Researchers retrospectively processed more than 40,000 2D mammograms using an AI cancer detection system (Transpara version 1.7.0, ScreenPoint Medical) to compare its detection rates to that of the original radiologists. Four years of follow-up data were reviewed for instances of eventual cancer diagnoses and factors associated with positive screenings. The group compared the results of double reading to that of AI as both a second reader and a standalone reader.
There were 580 cases deemed positive at follow-up. Double human reading recalled 1,244 mammograms, while single human reading combined with AI as a second reader resulted in 2,112 recalled mammograms. Use of AI improved sensitivity by approximately 8.4%.
Of the additional cancers detected by AI, 26.7% were considered invasive and 16.6% of those tumors had grown to 20 mm in diameter by the time they were spotted by radiologists at follow-up. The tool’s performance was consistent across different demographics and density categories, earning researchers’ support as a second reader.
“AI flags lesions that, according to a human radiologist, seemed of no significant clinical value at the time of screening, which occurs in all density categories. Part of these findings were later (sometimes much later) confirmed as malignant tumors, after a diagnosis due to symptoms or during the next screening round,” the team wrote. “Additionally, these tumors were, on average, more advanced at the time of diagnosis than the average screen-detected cancers. Thus, the AI-identified lesions that were unrecognized during human reading include clinically relevant cancers that progress when left untreated.”
Read more about the study’s findings here.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.