PGMI assessment in mammography: AI software versus human readers

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Assessing mammogram quality is crucial for effective breast cancer screening.
• Human reader variability in the Perfect-Good-Moderate-Inadequate (PGMI) classification can impact screening program quality.
• Exploring AI as a tool to standardize and improve image quality assessment is warranted.

Purpose of the Study:

• To evaluate human inter-reader agreement for PGMI classification of screening mammograms.
• To explore the role of artificial intelligence (AI) as an alternative reader for mammogram quality assessment.

Main Methods:

• Five radiographers from three European countries independently assessed 520 mammograms using PGMI.
• Dedicated AI software was used as a sixth reader.
• Accuracy, Cohen's Kappa, and confusion matrices compared AI and human reader performance.

Main Results:

• Significant inter-reader variability was observed among human readers (κ = -0.018 to 0.41).
• AI software surpassed human agreement in specific areas like glandular tissue cuts and pectoral angle measurement.
• AI performance was comparable to human assessment for other features and overall image quality.

Conclusions:

• High human inter-reader disagreement in PGMI assessment highlights a need for standardization.
• AI software shows potential for reliable, automated assessment of diagnostic image quality.
• AI can offer objective feedback to support radiographers in quality management, though workflow integration requires further attention.