From explanation to intervention: Interactive knowledge extraction from Convolutional Neural Networks used in...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning Interpretability
• Clinical Decision Support Systems

Background:

• Deep Learning models, particularly Convolutional Neural Networks (CNNs), excel at extracting complex features from medical X-ray images.
• Limited interpretability of CNNs hinders their adoption in clinical practice due to a lack of clinician trust.
• Bridging the gap between AI capabilities and clinical acceptance requires interpretable AI solutions.

Purpose of the Study:

• To develop an interactive framework to increase clinician trust in CNN-based medical image analysis.
• To enable clinicians to question and intervene in CNN decision-making processes.
• To create an interpretable refinement of data-driven CNNs aligned with medical best practices.

Main Methods:

• Translating a layer of a trained CNN into a compact set of symbolic rules.
• Utilizing expert interactions with rule visualizations to promote clinically-relevant CNN kernel usage.
• Employing radiomics analyses and permutation evaluations to define and validate kernel relevance.
• Removing clinically non-meaningful CNN kernels without compromising model performance.

Main Results:

• The framework successfully translates CNN layers into interpretable symbolic rules.
• Clinician interaction with rule visualizations identified and promoted clinically-relevant CNN kernels.
• Removal of non-meaningful kernels did not negatively impact overall model performance.
• The approach yielded an interpretable CNN refinement aligned with clinical expertise.

Conclusions:

• The proposed interactive framework enhances clinician trust by making CNN decisions transparent and controllable.
• Symbolic rule translation and expert-guided kernel refinement are effective strategies for interpretable AI in radiology.
• This method facilitates the integration of AI tools into clinical workflows by ensuring alignment with medical best practices.