← Back to Home

AI-assisted mammography interpretation improves breast cancer lesion detection in retrospective study

In imaging data from 321 patients with breast cancer, combined interpretation by physicians and AI achieved a lesion detection rate of 90.6%, 3 percentage points higher than physicians alone. However, all the data came from patients with cancer, leaving unanswered how many additional false positives might occur in general screening.

By SURL BioNews

Detecting one more cancer lesion on a mammogram may provide another clue for diagnosis and treatment. However, whether increased detection also leads to more examinations for people without cancer likewise affects a tool’s clinical value. A new study shows that combined interpretation by artificial intelligence (AI) and radiologists can improve lesion detection in patients with confirmed breast cancer, but the answer regarding its overall benefit in general screening remains incomplete.

The peer-reviewed study, published on October 23 in the *Journal of Imaging Informatics in Medicine*, retrospectively analyzed 331 lesions in 321 patients with breast cancer, comparing interpretation by physicians alone, AI alone, and the two combined. It assessed image recognition using existing cases and was not a prospective screening trial in a general screening population.

The study reported a lesion detection rate of 90.6% for combined interpretation by physicians and AI, compared with 87.6% for physicians alone, an absolute difference of 3 percentage points. The denominator here is the cancer lesions included in the study. This cannot be interpreted as diagnostic accuracy among everyone undergoing mammography, nor does it mean there is already evidence of a reduction in the risk of death.

The AI system used in the study outputs Breast Imaging Reporting and Data System (BI-RADS) categories, presenting the algorithm’s results in a classification format familiar to clinicians. The study also compared human and AI assessments of imaging features and found that agreement varied by feature. This suggests that the value of combined interpretation may come from complementary recognition capabilities, with the degree of assistance varying across types of imaging findings.

The most important limitation is that all participants had breast cancer. Without people who do not have cancer, specificity—the ability to correctly rule out cancer in noncancer cases—cannot be fully assessed. Likewise, these findings cannot establish how much adding AI would increase or decrease false positives or the resulting burden of additional imaging examinations or biopsies.

These results provide limited but concrete support for AI-assisted interpretation: in this set of cancer cases, combined human and AI interpretation did detect a higher proportion of lesions. To determine whether it can improve routine screening, further validation must include people without cancer and assess missed cancers, false positives, and additional examinations together in actual workflows. Only then can the overall benefit of detecting more lesions become clear.

References

  1. Journal of Imaging Informatics in Medicine