Lunit Study in Radiology Highlights Trust Gap Between Radiologists and AI in Breast Cancer Screening
Posted on April 17, 2025 17:14
Modified on April 17, 2025 17:26
Despite strong performance from AI, fewer recalls from AI-flagged cases point to challenges in human-AI collaboration
[Press Release] SEOUL, South Korea, April 9, 2025-- Lunit(루닛), a leading provider of AI-powered solutions for cancer diagnostics and therapeutics, today announced the publication of a new study in Radiology (RSNA's flagship journal), highlighting how radiologists interact with AI in real-world breast cancer screening.
Lunit’s AI identified more cancers, but radiologists recalled fewer of those cases—highlighting a trust gap in clinical decision-making.
The study, based on the prospective ScreenTrustCAD trial, is the first large-scale, population-based trial to evaluate how radiologists interact with AI in an operational breast cancer screening setting. Conducted at Capio St Göran Hospital in Stockholm, Sweden, the trial enrolled about 55,000 women and implemented Lunit INSIGHT MMG as an independent third reader alongside two radiologists.
The findings suggest that radiologists were significantly less likely to recall patients flagged by AI—even though those cases were more likely to be cancerous—pointing to a trust gap between human clinicians and AI tools.
"This study isn't just about how well the AI performs—it's about what happens when AI enters the decision-making process," said Brandon Suh, CEO of Lunit. "The findings point to a gap between AI input and human response, and that's where real-world impact is made or lost."
Human-AI Collaboration
The study showed that:
-When only AI flagged a case, just 4.6% were recalled by radiologists.
-When only a radiologist flagged a case, 14.2% were recalled.
-When two radiologists flagged a case, 57.2% were recalled—but when AI and one radiologist both flagged it, the recall rate dropped to 38.6%. Despite being blinded to AI scores during the initial review, radiologists consistently recalled fewer patients from AI-flagged cases—raising questions about how AI findings are weighed in clinical judgment.
"This isn't a question of whether AI can detect cancer. It's about how AI findings are interpreted and acted on by the people making clinical decisions," said Dr. Karin Dembrower, lead author of the study and radiologist at Karolinska Institutet.
Clinical Performance of Lunit INSIGHT MMG
While the study's central focus was on decision-making dynamics, it also confirmed the strong diagnostic performance of Lunit INSIGHT MMG:
-Among women recalled after screening, 22% of AI-only flagged cases were diagnosed with cancer.
-In contrast, just 3.4% of single-radiologist flags and 2.5% of two-radiologist flags resulted in a cancer diagnosis.
-When both AI and one radiologist flagged a case, the cancer yield was highest at 25%. These results highlight AI's ability to detect cancers that may go unnoticed in standard double-reading workflows—particularly in early or less obvious presentations.
These findings suggest that while the AI system correctly identified high-risk cases, many of those cases were not recalled by radiologists—indicating a potential gap in trust or confidence in AI-generated findings.
"The AI clearly shows strong diagnostic performance," Suh added. "What this study reveals is that realizing that potential in daily clinical practice depends on how AI and humans work together."
News provided by Lunit
[Press Release] SEOUL, South Korea, April 9, 2025-- Lunit(루닛), a leading provider of AI-powered solutions for cancer diagnostics and therapeutics, today announced the publication of a new study in Radiology (RSNA's flagship journal), highlighting how radiologists interact with AI in real-world breast cancer screening.
Lunit’s AI identified more cancers, but radiologists recalled fewer of those cases—highlighting a trust gap in clinical decision-making.
The study, based on the prospective ScreenTrustCAD trial, is the first large-scale, population-based trial to evaluate how radiologists interact with AI in an operational breast cancer screening setting. Conducted at Capio St Göran Hospital in Stockholm, Sweden, the trial enrolled about 55,000 women and implemented Lunit INSIGHT MMG as an independent third reader alongside two radiologists.
The findings suggest that radiologists were significantly less likely to recall patients flagged by AI—even though those cases were more likely to be cancerous—pointing to a trust gap between human clinicians and AI tools.
"This study isn't just about how well the AI performs—it's about what happens when AI enters the decision-making process," said Brandon Suh, CEO of Lunit. "The findings point to a gap between AI input and human response, and that's where real-world impact is made or lost."
Human-AI Collaboration
The study showed that:
-When only AI flagged a case, just 4.6% were recalled by radiologists.
-When only a radiologist flagged a case, 14.2% were recalled.
-When two radiologists flagged a case, 57.2% were recalled—but when AI and one radiologist both flagged it, the recall rate dropped to 38.6%. Despite being blinded to AI scores during the initial review, radiologists consistently recalled fewer patients from AI-flagged cases—raising questions about how AI findings are weighed in clinical judgment.
"This isn't a question of whether AI can detect cancer. It's about how AI findings are interpreted and acted on by the people making clinical decisions," said Dr. Karin Dembrower, lead author of the study and radiologist at Karolinska Institutet.
Clinical Performance of Lunit INSIGHT MMG
While the study's central focus was on decision-making dynamics, it also confirmed the strong diagnostic performance of Lunit INSIGHT MMG:
-Among women recalled after screening, 22% of AI-only flagged cases were diagnosed with cancer.
-In contrast, just 3.4% of single-radiologist flags and 2.5% of two-radiologist flags resulted in a cancer diagnosis.
-When both AI and one radiologist flagged a case, the cancer yield was highest at 25%. These results highlight AI's ability to detect cancers that may go unnoticed in standard double-reading workflows—particularly in early or less obvious presentations.
These findings suggest that while the AI system correctly identified high-risk cases, many of those cases were not recalled by radiologists—indicating a potential gap in trust or confidence in AI-generated findings.
"The AI clearly shows strong diagnostic performance," Suh added. "What this study reveals is that realizing that potential in daily clinical practice depends on how AI and humans work together."
News provided by Lunit