Comparing Radiologists' and Artificial Intelligence Performance Detecting Suspicious Microcalcifications on Screening...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Breast Cancer Screening

Background:

• Early breast cancer detection via mammography is vital for reducing mortality.
• Microcalcifications are common mammographic findings with varied interpretations.
• Accurate identification of suspicious calcifications is critical for patient recall and diagnosis.

Purpose of the Study:

• To evaluate radiologist performance in identifying suspicious calcifications on screening mammograms.
• To compare human reader performance against an in-house artificial intelligence (AI) model.
• To assess the impact of experience and caseload on reader accuracy.

Main Methods:

• A cross-sectional pilot study involving 27 radiologists, 2 breast physicians, and 6 trainees.
• A test set of 30 mammographic cases (10 malignant, 20 benign/normal) with diverse calcifications.
• Comparison of reader performance to the Sydney-GMIC AI model using statistical tests.

Main Results:

• Radiologists with ≤10 years of experience showed higher sensitivity than trainees (72.2% vs 53.3%).
• Higher weekly caseloads correlated with decreased specificity but increased sensitivity.
• The Sydney-GMIC AI model outperformed the average radiologist in both sensitivity and specificity.

Conclusions:

• Recalling calcifications from screening mammograms presents challenges with variable human reader performance.
• AI models show potential utility in mammographic analysis for screening cases.
• AI performance suggests a complementary role in improving diagnostic accuracy for calcification assessment.