One of the most exciting radiology use cases for AI is in breast screening. At last week’s RSNA 2023 show, a paper highlighted the technology’s potential for helping breast imagers focus on cases more likely to have cancer.
Looking for cancers on screening mammography has been compared to finding a needle in a haystack, and as such it’s considered to be one of the areas where AI can best help.
- One of the earliest use cases was in identifying suspicious breast lesions during radiologist interpretation (remember computer-aided detection?), but more recently researchers have focused on using AI as a triage tool, by identifying cases most likely to be normal that could be removed from the radiologist’s urgent worklist. Studies have found that 30-40% of breast screening cases could be read by AI alone or triaged to a low-suspicion list.
But what impact would AI-based breast screening triage have on radiologist metrics such as recall rate?
- To answer this question, researchers from NYU Langone Health prospectively tested their homegrown AI algorithm for analyzing DBT screening cases.
The algorithm was trained to identify extremely low-risk cases that could be triaged from the worklist while more complex cases where the AI was uncertain were sent to radiologists, who knew in advance the cases they were reading were more complicated. In 11.7k screening mammograms, researchers examined recall rates over two periods, one before AI triage and one after, finding:
- The overall recall rate went from 13% before the triage period to 15% after
- Recall rates for complex cases went from 17% to 20%
- Recall rates for extremely low-risk studies went from 6% to 5%
- There were no statistically significant differences in any of the comparisons
- No change in median self-reported perceived difficulty of reading from the triage lists compared to non-triage list, regardless of years of experience
In future work, the NYU Langone researchers will continue their study to look at AI’s impact on cancer detection rate, biopsy rate, positive predictive value, and other metrics.
The NYU Langone study puts a US spin on research like MASAI from Sweden, in which AI was able to reduce radiologists’ breast screening workload by 44%. Given the differences in screening protocols between the US and Europe, it’s important to assess how AI affects workload between the regions.
Further work is needed in this ongoing study, but early results indicate that AI can triage complex cases without having an undue impact on recall rate or self-perceived difficulty in interpreting exams – a surrogate measure for burnout.