Assessment of breast positioning criteria in mammographic screening: Agreement between artificial intelligence software and radiographers

Objectives To determine the agreement between artificial intelligence software (AI) and radiographers in assessing breast positioning criteria for mammograms from standard digital mammography and digital breast tomosynthesis. Methods Assessment of breast positioning was performed by AI and by four radiographers in pairs of two on 156 examinations of women screened in Bergen, April to September 2019, as part of BreastScreen Norway. Ten criteria were used; three for craniocaudal and seven for mediolateral-oblique view. The criteria evaluated the appearance of the nipple, breast rotation, pectoral muscle, inframammary fold and pectoral nipple line. Intraclass correlation and Cohen's kappa coefficient (kappa) were used to investigate the correlation and agreement between the radiographer's assessments and AI. Results The intraclass correlation for the pectoral nipple line between the radiographers and AI was >0.92. A substantial to almost perfect agreement (kappa > 0.69) was observed between the radiographers and AI on the nipple in profile criterion. We observed a slight to moderate agreement for the other criteria (kappa = 0.06-0.52) and generally a higher agreement between the two pairs of radiographers (mean kappa = 0.70) than between the radiographers and AI (mean kappa = 0.41). Conclusions AI has great potential in evaluating breast position criteria in mammography by reducing subjectivity. However, varying agreement between radiographers and AI was observed. Standardized and evidence-based criteria for definitions, understandings and assessment methods are needed to reach optimal image quality in mammography.