Radiomic analysis of cohort-specific diagnostic errors in reading dense mammograms using artificial intelligence.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Radiologists face challenges interpreting dense screening mammograms.
• Accurate interpretation is crucial for early breast cancer detection.

Purpose of the Study:

• To investigate radiomics-based artificial intelligence (AI) for identifying radiologists' interpretation errors in dense mammograms.
• To analyze distinct radiomic features associated with false positives and false negatives across different reader cohorts.

Main Methods:

• Thirty-six radiologists from China and Australia interpreted 60 dense mammograms.
• Radiomic features were extracted from suspicious and malignant areas.
• Random forest models were trained to identify error-prone regions.

Main Results:

• The AI model showed varying performance across cohorts and views (e.g., AUC for false positives in Chinese cohort: 0.864 CC, 0.829 MLO).
• Distinct radiomic features, like Gabor and maximum response filter outputs, were linked to false positives.
• Intensity changes and coarse textures correlated with false negatives.

Conclusions:

• A radiomics-based AI pipeline effectively identifies common interpretation errors in specific radiologist cohorts.
• This approach highlights distinct image features contributing to diagnostic errors in dense mammography interpretation.