Interobserver agreement between radiologists and artificial intelligence in mammographic breast density classification

Abstract

Artificial intelligence (AI) has been proposed as a tool for assessing mammographic breast density (MBD). This study aimed to evaluate the agreement of MBD classification between four radiologists (human readers [HRs]) with different years of experience in breast imaging and the AI Lunit INSIGHT MMG. This cross-sectional study was conducted with a convenience sample of radiologists trained in breast imaging who assessed MBD screening mammograms of asymptomatic women 35 years or older using BI-RADS descriptors. Cohen’s kappa determined the agreement between the HRs and AI. A total of 192 women with a mean age of 55.4 ± 31.8 years (range 37-82 years) were included. Interobserver agreement between HRs and AI varied in Category a but was substantial in Category b (HR1 k = 0.729, HR2 k = 0.718, HR3 k = 0.768, and HR4 k = 0.672) and in Category c, HR1, HR2, and HR3 had substantial agreement (k = 0.728, k = 0.697, and k = 0.738, respectively) and HR4 had moderate agreement (k = 0.578), while in Category d, it was mostly moderate. HRs and AI agreements varied from fair to substantial. HRs with more years of experience in breast image interpretation had a lower agreement with AI for MBD classification than HRs with less time.