Artificial Intelligence Using Open Source BI-RADS Data Exemplifying Potential Future Use.

Area of Science:

• Radiology
• Artificial Intelligence
• Machine Learning

Background:

• Concerns exist regarding artificial intelligence (AI) potentially replacing radiologists.
• Evaluating a radiologist-augmented AI workflow addresses this by integrating human expertise with AI.
• The study hypothesizes that combining radiologist input with AI improves predictive accuracy over AI alone.

Purpose of the Study:

• To assess if incorporating radiologist-provided BI-RADS classifications improves AI prediction accuracy.
• To compare the performance of machine learning algorithms with and without radiologist input.
• To evaluate the efficacy of a collaborative AI-radiologist approach in medical diagnosis.

Main Methods:

• Utilized open-source BI-RADS datasets (UCI Machine Learning Repository and Digital Database for Screening Mammography).
• Trained three machine learning algorithms using 10-fold cross-validation.
• Developed two model sets: M1 (image parameters only) and M2 (image parameters plus BI-RADS classification).

Main Results:

• Models incorporating radiologist-provided BI-RADS classification demonstrated significantly superior performance.
• Statistical analysis showed a P < .0001 for the improved accuracy of M2 models.
• The area under the curve and Gini coefficient were higher for models including radiologist input.

Conclusions:

• AI and radiologists working collaboratively achieve superior diagnostic outcomes.
• This integrated approach aids in case-based decision-making for improved patient care.
• Further research into AI predictor metrics can yield deeper insights into medical imaging analysis.