By Mary Henderson
Results of the largest retrospective clinical study in breast screening showed that the combination of human readers and AI has the potential to improve screening performance, especially among women with dense breasts. Annie Ng, PhD, presented the study’s results during a Monday morning session.
“While women with dense breasts have a higher probability of developing breast cancer, it’s harder to see in dense breast tissue, so they’re more likely to be missed,” said Dr. Ng, science lead at Kheiron Medical Technologies, Ltd., a part of DeepHealth.
Dr. Ng said AI complements human readers by bringing a different visual assessment to the mammogram reading, reducing bias in the single-read and eliminating redundancy in the double-read.
“The strengths and weaknesses of the human reader and AI actually complement each other,” she said. “By combining the different visual perceptions of AI and human reading, you widen the scope of capability and maximize the effectiveness of finding cancers while reducing false positives.”
Dr. Ng’s research team compared the performance of a commercial AI tool to standard human reading for detecting cancer on screening mammography across different density categories. The team used breast screening data from three sites in the UK, which included three major genetic clusters and diverse ethnicities in London.
The cohort included 306,839 screening cases from 236,739 participants (age 49-71) performed between 2017 and 2021. Breast density was determined according to the four BI-RADS categories, in which A or B is defined as fatty, and C or D is defined as dense. Of the subjects, 58.2% had fatty breasts and 41.8% had dense breasts.
Within the cohort, there was a total of 2,965 pathology-proven cancer cases, including 2,588 screen-detected (SD) cancers and 379 interval cancers (IC).
In the first comparison, the researchers assessed the sensitivity rates of AI and a single human reader. AI outperformed across all cases and for dense breasts. This finding held, regardless of patient age.
“Human readers had a non-significant higher sensitivity than the AI system for cancer detection in fatty breasts, but the effect was opposite in dense breasts, where AI showed a significantly higher sensitivity than human readers,” she said.
Dr. Ng recommended using AI in two different roles to best complement human reading: as an independent reader that serves as a second reader when in agreement with the first; and as an extra reader, providing an additional review for non-recall cases.
The researchers secondly modeled a double reading with AI in the supporting independent reader role and compared it to the human double reading.
Double reading with AI had similar cancer detection rates but a lower recall rate for fatty and dense breasts and a higher positive predictive value than the standard human double reading. According to the model, using AI as an independent reader would result in a substantial human reading workload savings of -42.5%.
“The use of AI as both an independent and extra reader may achieve the best overall screening outcomes for equitable breast screening,” Dr. Ng concluded.
Access the presentation, “Improving Cancer Detection in Dense Breasts in Breast Screening Programs: The Complementary Strength of AI and Human Intelligence,” (M1-SSBR03-2) on demand at RSNA.org/MeetingCentral.
© 2024 RSNA.
The RSNA 2024 Daily Bulletin is the official publication of the 110th Scientific Assembly and Annual Meeting of the Radiological Society of North America. Published online Sunday, December 1 — Friday, December 6.
The RSNA 2024 Daily Bulletin is owned and published by the Radiological Society of North America, Inc., 820 Jorie Blvd., Suite 200, Oak Brook, IL 60523.
