Performance benchmarking

ECR 2019

28 February 2019


Alejandro Rodriguez-Ruiz et al

Artificial intelligence detecting breast cancer on mammography: does breast density play a role?

Alejandro Rodriguez-Ruiz, Michiel Kallenberg, Albert Gubern-Merida, Nico Karssemeijer, Ritse M. Mann


To analyze the relationship between the breast cancer detection performance of an artificial intelligence (AI) system on mammography and breast density.

Materials and Methods

An independent multi-vendor cancer-enriched database of 1397 mammograms was collected across several institutions in Europe (179 with biopsy-verified cancer, 1218 normal with at least two years of negative follow-up). For each mammogram, an AI system computed a score representing the likelihood of presence of cancer (scale 1-10); while the breast density volume fraction (averaged across views) was computed with previously published and validated software. Linear regression analysis was performed for AI score and density. Subsequently, the population was divided into two groups based on the median density value of the population: high-density mammograms (73 cancers, 626 normals) and low-density mammograms (106 cancers, 592 normals). The area under the receiver operating characteristic curve (AUC) of the AI system was compared between the two groups.


Median density volume fraction in the population was 0.12 (0.09-0.18). The AI score did not correlate with density (R2 = -0.09), with low AI score categories harboring at same proportions high-density and low-density mammograms. The AUC of the AI system was similar for both groups (0.91 vs 0.89, difference = +0.02, 95% CI = -0.03,0.08).


The AI score may be considered as an independent tool to estimate the likelihood of the presence of cancer on mammograms, to stratify screening populations, and to potentially fasten the reading process by reassuring readers on mammograms that are likely normal, irrespective of breast density.

You might also like