Enhancing interpretability of automatically extracted machine learning features: application to a RBM-Random Forest...

Area of Science:

• Medical Imaging Analysis
• Machine Learning
• Artificial Intelligence in Healthcare

Background:

• Machine learning models offer high performance but suffer from complexity and lack of interpretability, hindering trust in critical medical applications.
• Representation learning techniques, while effective for automatic feature extraction, are often treated as "black boxes", limiting their clinical utility.
• Interpretability is crucial for the adoption of AI in medicine, enabling validation and trust in automated diagnostic and prognostic tools.

Purpose of the Study:

• To propose and evaluate a methodology for enhancing the interpretability of automatically extracted machine learning features in medical imaging.
• To develop a system combining unsupervised feature learning with a classifier to analyze correlations between imaging data, features, and clinical outcomes.
• To provide both global and local interpretation levels for machine learning models applied to medical tasks.

Main Methods:

• A hybrid system integrating a Restricted Boltzmann Machine for unsupervised feature learning and a Random Forest classifier.
• Development of a novel feature importance strategy considering both imaging data and target variables.
• Evaluation of the methodology on two clinical tasks: brain tumor segmentation and penumbra estimation in ischemic stroke.

Main Results:

• The proposed methodology successfully enhances the interpretability of learned features in both brain tumor segmentation and ischemic stroke lesion analysis.
• Demonstrated ability to reveal relationships between different imaging modalities and extracted features, crucial for clinical tasks.
• Identified specific learning patterns that mimic expert analysis of medical images, validating the approach's clinical relevance.

Conclusions:

• The developed methodology significantly improves the interpretability of complex machine learning models in medical imaging.
• The approach provides valuable insights into feature relevance and model decision-making processes at both global and local levels.
• This work contributes to building trust and facilitating the adoption of interpretable AI in critical healthcare domains.