AI for reading screening mammograms: the need for circumspection

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Current studies evaluating artificial intelligence (AI) for screening mammogram interpretation face significant methodological limitations.
• These limitations challenge the purported superiority of AI over experienced radiologists in detecting breast cancer.

Discussion:

• The existing research does not adequately address the potential for overdiagnosis, failing to quantify additional breast cancers identified by AI.
• A critical gap exists in evaluating the impact of AI on subsequent diagnostic procedures, such as biopsies, for positive mammogram readings.

Key Insights:

• Methodological flaws in AI mammogram studies cast doubt on AI's performance compared to radiologists.
• The risk of overdiagnosis due to AI detection of clinically insignificant cancers remains unquantified.
• The absence of data on biopsy recommendations for AI-detected abnormalities inflates AI's perceived accuracy.

Outlook:

• Future research must incorporate robust methodologies to accurately assess AI performance in mammography.
• Studies need to investigate the clinical significance of AI-detected findings and their impact on patient outcomes.
• Standardized reporting on biopsy rates following AI interpretation is crucial for a realistic evaluation of AI in breast cancer screening.