Comparing percent breast density assessments of an AI-based method with expert reader estimates: inter-observer...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Breast density is a key factor in breast cancer risk assessment.
• Manual assessment of mammographic density by experts shows significant inter-observer variability.
• Automated methods are being developed to improve consistency and accuracy in breast density estimation.

Purpose of the Study:

• To investigate the inter-reader variability between expert assessors and a deep learning approach for breast density estimation.
• To compare the risk prediction capabilities of expert readers and an automated deep learning model.

Main Methods:

• Utilized screening data from a cohort of 1328 women.
• Compared two expert readers and a single reader against the Manchester artificial intelligence - visual analog scale (MAI-VAS) deep learning model.
• Employed Bland-Altman analysis for variability assessment and matched concordance index for risk prediction.

Main Results:

• The MAI-VAS deep learning model demonstrated substantially lower limits of agreement (SD ±21) compared to two expert readers (SD ±31).
• Inter-observer agreement for the AI tool with a single expert was comparable to that between two experts.
• Breast cancer risk discrimination by the deep learning method was similar to that of a single expert (concordance 0.628 vs. 0.624).

Conclusions:

• The artificial intelligence breast density assessment tool MAI-VAS exhibits superior inter-observer agreement compared to the agreement between two human experts.
• Deep learning-based methods for breast density assessment offer consistent scores without compromising breast cancer risk prediction.