A framework for falsifiable explanations of machine learning models with an application in computational pathology

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Machine learning explainability

Background:

• Deep learning drives advancements in medical diagnosis and treatment via image analysis.
• Neural networks often function as "black boxes," lacking transparency in their decision-making processes.
• Formalizing explainability for these models remains a significant challenge.

Purpose of the Study:

• To introduce a novel hypothesis-based framework for creating falsifiable explanations of machine learning models.
• To address the "black box" problem in artificial intelligence for medical applications.
• To enhance the transparency and reliability of AI-driven diagnostic tools.

Main Methods:

• Developed a hypothesis-based framework for falsifiable explanations.
• Defined falsifiable explanations as hypotheses linking model-induced intermediate spaces to data origins.
• Applied the framework to computational pathology using hyperspectral infrared microscopy.
• Used activation maps trained with an inductive bias for tumor localization.
• Validated explanations by hypothesizing activation-image correspondence and confirming with histological staining.

Main Results:

• Successfully instantiated a falsifiable explanation framework in a computational pathology context.
• Demonstrated that activation maps can be hypothesized to correspond to tumor and associated structures.
• Validated these hypotheses using independent histological staining, confirming the framework's efficacy.
• Provided a method for making AI explanations in medical imaging more transparent and testable.

Conclusions:

• The proposed framework offers a method for generating falsifiable explanations for machine learning models.
• This approach enhances transparency in computational pathology by linking AI outputs to biological realities.
• Falsifiable explanations, validated through independent experiments, can improve trust and reliability in AI-assisted medical diagnosis.