On the interplay of machine learning and background knowledge in image interpretation by Bayesian networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biostatistics

Background:

• Bayesian networks are powerful tools for medical image interpretation.
• Expert knowledge and data-driven learning methods are crucial for building effective Bayesian networks.
• Mammography interpretation presents unique challenges due to continuous image features.

Purpose of the Study:

• To evaluate the optimal balance between expert knowledge and machine learning in constructing Bayesian networks for medical image interpretation.
• To compare different methods for Bayesian network construction using mammogram data.
• To assess the impact of data discretization and structure learning on diagnostic performance.

Main Methods:

• Mammogram interpretation was used as a case study.
• Two Bayesian network construction methods were explored: discretization of continuous features and structure learning from data.
• Gaussian and multinomial distributions were compared for probabilistic parameters.
• Structure learning was performed on screening mammographic data from 795 patients.

Main Results:

• Discretized data in Bayesian networks improved cancer detection by up to 11.7% and offered better interpretability compared to expert-based Gaussian models.
• Structure learning revealed novel relationships between mammographic features, complementing expert-derived knowledge.
• Combining discretized features and structure learning enhanced cancer detection by up to 17% over manual models.

Conclusions:

• A balanced approach integrating expert and data-derived knowledge is essential for effective Bayesian network construction in medical image interpretation.
• Optimizing network structure and parameters through a hybrid approach leads to more accurate and understandable diagnostic models.