Mammography breast density assessment with Quantra artificial intelligence software: intra- and interobserver agreement with radiologists

Abstract

Introduction. A new version of Quantra v.2.2.2 artificial intelligence (AI) software is used for mammography breast density (MBD) assessment. This study aimed to determine the intraobserver agreement of radiologists (human readers, HRs) and the interobserver agreement between AI Quantra and HRs in MBD assessment. Material and methods: This prospective cohort study was conducted in two phases. HRs and AI Quantra version 2.2.2 assessed MBD in 2D mammograms and tomosynthesis in women over 35 years based on the BI-RADS 5th Edition. HRs had training in breast imaging with different levels of experience. MBD assessment between the first (phase 1) and second reading (phase 2) of the HRs was compared using four (a, b, c, d) and two MBD categories (a+b, non-dense, and c+d, dense). Intraobserver and interobserver agreement was assessed with light Kappa and Cohen's Kappa. Results: Four HRs with 2 to 32 years of experience interpreting breast images participated. They evaluated 685 mammograms in each phase. Intraobserver agreement between the two phases was moderate to substantial for the four MBD categories and substantial to near perfect for the two MBD categories. The interobserver agreement between AI Quantra and the four HRs varied from slight to moderate for the four MBD categories and moderate to substantial for the two categories. The interobserver agreement between AI and the four HRs improved in phase 2. Conclusion: Intraobserver agreement between radiologists and interobserver agreement between AI Quantra and HRs increased when assessing non-dense and dense MBD categories. Many factors can influence the visual assessment of MBD by radiologists, making AI a tool that can provide radiologists with additional information when assessing MBD.