Computer-assisted diagnosis of breast cancer using a data-driven Bayesian belief network

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Biomedical Data Analysis

Background:

• Accurate breast cancer diagnosis relies on integrating diverse data sources.
• Current diagnostic models often use hybrid approaches, combining separate analyses of image and non-image data.
• Evaluating the efficacy of a unified network for simultaneous feature integration is crucial.

Purpose of the Study:

• To investigate a Bayesian belief network for breast cancer diagnosis.
• To compare the diagnostic performance of a single integrated network versus hybrid networks.
• To determine if combining image and non-image features in one network improves accuracy.

Main Methods:

• Developed three Bayesian belief networks using data from 419 breast cancer cases (92 malignancies).
• Extracted 13 features including mammographic findings, physical exams, and clinical history.
• Compared performance using average area under the ROC curve (Az) for image-only, non-image-only, hybrid, and integrated networks.

Main Results:

• The integrated network incorporating all features achieved the highest diagnostic performance (Az = 0.89).
• Hybrid networks (Az = 0.85-0.87) outperformed networks using only image (Az = 0.81) or non-image (Az = 0.71) features.
• A single network simultaneously evaluating all information sources demonstrated superior accuracy.

Conclusions:

• A single Bayesian belief network integrating both image and non-image features offers superior diagnostic performance for breast cancer compared to hybrid approaches.
• This unified approach mimics human observers by concurrently processing diverse information types.
• Further research is warranted to validate these findings in larger datasets.