Using AI to Identify Errors in Reading Dense Mammograms
Author Information
Author(s): Tao Xuetong, Gandomkar Ziba, Li Tong, Brennan Patrick C, Reed Warren M
Primary Institution: The University of Sydney
Hypothesis
Can a radiomics-based AI approach effectively identify diagnostic errors made by radiologists when reading dense mammograms?
Conclusion
The study successfully identified common diagnostic errors among radiologists reading dense mammograms using AI, highlighting specific features linked to false positives and negatives.
Supporting Evidence
- The AI model showed distinct patterns of errors among radiologists from different geographic cohorts.
- Radiologists with similar geographic characteristics exhibited comparable error patterns.
- Specific radiomic features were linked to higher rates of false positives and negatives.
Takeaway
This study shows that AI can help find mistakes radiologists make when looking at dense mammograms, which can help them improve.
Methodology
Thirty-six radiologists from China and Australia read 60 dense mammograms, and radiomic features were extracted to train random forest models to identify diagnostic errors.
Potential Biases
Potential bias due to the familiarity of Australian radiologists with the test set.
Limitations
The study used mammograms from the Australian population, which may not fully represent the Chinese cohort's experience.
Participant Demographics
36 radiologists from China (20) and Australia (16) with varying levels of experience.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website