AI cancer detection on screening mammography performs well but accuracy falls as breast density increases

In 200,000 screening exams, an AI model's AUC for breast cancer detection was highest in the least dense breasts (0.955) and lowest in the densest (0.857). Density distribution varied by vendor, and 18.5% of exams had different density categories for the right and left breast.

• The study retrospectively analyzed 200,000 exams from BreastScreen Norway using a commercial AI tool that provides automated volumetric breast density (VBD) and a malignancy risk score.
• AUC for cancer detection differed significantly across density groups: VBD 1 (least dense) AUC 0.955 (95% CI 0.937-0.992) versus VBD 4 (most dense) AUC 0.857 (95% CI 0.803-0.910).
• Density classification was vendor-dependent: 3.3% of exams from vendor A were categorized as VBD 4 versus 6.7% from vendor B, and 18.5% of exams had discordant VBD categories between right and left breasts.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.