por John R. Fischer
, Senior Reporter | April 15, 2022
Based on its ability to detect cancer, artificial intelligence may serve as a sufficient second reader for mammograms and reduce the workload on radiologists.
That’s what researchers in Norway are saying in a new study comparing the technology’s performance to routine independent double reading in a population-based screening program.
The largest of its kind, the study assessed the use of AI in reading almost 123,000 exams performed on over 47,000 women from four facilities in BreastScreen Norway, the country’s population-based screening programs. Such programs conduct so many mammograms that produce significant workloads for radiologists and can lead to backlogs and longer waiting times for patients. And while AI has shown encouraging results in identifying cancer, its use in real screening settings is limited.
Using a commercially available AI system, the researchers were able to identify and eliminate a high percentage of benign exams from their workload, as well as find the majority of screen-detected cancers. Less than 20% of screen-detected cancers were not identified.
The researchers say these findings show the potential that AI has for reducing interpretive volumes. "Based on our results, we expect AI to be of great value in the interpretation of screening mammograms in the future. We expect the greatest potential to be in reducing the reading volume by selecting negative examinations,” said Solveig Hofvind, from the Section for Breast Cancer Screening, Cancer Registry of Norway in Oslo, who led the study.
Using a scale of one to 10 (one being the lowest, 10 the highest), the system predicted the risk of cancer for each exam. The total number of cancers detected at screening was 752, along with 205 interval cancers found between screening rounds. The system assigned a score of 10 to 87.5% (653) of screen-detected cancers and to 44.9% (92) of interval cancers.
And when using AI as one of two readers in a double reading setting, the radiologist was still able to identify the small number of cancers that went undetected.
The researchers say the results show favorable histopathologic characteristics that are associated with a better prognosis for screen-detected cancers with low versus high AI scores. The opposite was found with interval cancers, which may imply that such cases with low scores are true interval cancers that are not visible on screening mammograms.
The researchers plan to conduct more prospective studies on the clinical use of AI in breast cancer settings with retrospective data but say this most recent one will serve as a basis for them and other future research.
"In our study, we assumed that all cancer cases selected by the AI system were detected," Dr. Hofvind said. "This might not be true in a real screening setting. However, given that assumption, AI will probably be of great value in interpretation of screening mammograms in the future."
The findings were published in Radiology