# A practical work around for breast density distribution discrepancies between mammographic images from different vendors

## TL;DR

This paper shows that different mammography machines can lead to different breast density readings, and proposes a method to ensure fair selection of high-risk women for breast cancer screening.

## Contribution

A novel percentile-matching method is proposed to mitigate vendor-specific biases in breast density evaluation.

## Key findings

- Breast density distributions differ significantly between devices from different vendors.
- A percentile-based approach ensures unbiased selection of high-risk women across different mammography systems.

## Abstract

Investigate the impact of mammography device grouped by vendor on volumetric breast density and propose a method that mitigates biases when determining the proportion of high-density women.

Density grade class and volumetric breast density distributions were obtained from mammographic images from three different vendor devices in different centers using breast density evaluation software in a retrospective study. Density distributions were compared across devices with a Mann–Whitney U test and breast density thresholds corresponding to distribution percentiles calculated. A method of matching density percentiles is proposed to determine women at potentially high risk while mitigating possible bias due to the device used for screening.

2083 (mean age 59 ± 5.4), 531 (mean age 58.8 ± 5.7) and 244 (mean age 60.7 ± 6.0) screened women were evaluated on three vendor devices, respectively. Both the density grade distribution and the volumetric breast density were different between Vendor 1 and Vendor 2 data (p < 0.001) and between Vendor 1 and Vendor 3 data (p < 0.001). Between Vendor 2 and Vendor 3, no significant difference was observed (p = 0.67 for density grade, p = 0.29 for volumetric density). To recruit the top 10% of women with extremely dense breasts required respective density thresholds of 16.1%, 13.6% and 13.8% for the three vendor devices.

Density grade class and volumetric breast density distributions differ between devices grouped by vendor and can result in statistically different breast density distributions. Percentile-dependent density thresholds can ensure unbiased selection of high-risk women.

Question
Does the use of x-ray systems from different vendors influence breast density evaluation and the resulting selection of high-risk women during breast cancer screening?

Findings
Statistically significant differences were observed between breast density distributions of different vendors; a method of matching via percentiles is proposed to prevent biased density evaluations.

Clinical relevance
Measured breast density distributions differed between X-ray devices. A workaround is proposed that determines density thresholds corresponding to a specified population, allowing the same proportion of women to be selected with a density algorithm.