Improving explanations for medical X-ray diagnosis combining variational autoencoders and adversarial machine...

Area of Science:

• Medical Computer Vision
• Artificial Intelligence in Healthcare
• Explainable AI

Background:

• Explainability is crucial for AI adoption in healthcare.
• Current AI models for medical diagnostics often lack transparency.

Purpose of the Study:

• To propose a novel Deep Learning architecture for explainable AI in medical diagnostics.
• To develop a method for generating visual explanations of AI decisions in medical imaging.

Main Methods:

• Leveraging Variational Autoencoders for image modification in a latent space.
• Utilizing a multi-objective genetic algorithm for explanation searching.
• Applying global and local regularization of the latent space.

Main Results:

• The proposed approach generates non-linear explanations in the original image space.
• The genetic algorithm efficiently searches for explanations without hyperparameter tuning.
• Achieved improved explanation precision by 56.39 to 7.23 percentage points compared to state-of-the-art methods.

Conclusions:

• The novel Deep Learning architecture offers effective explainability for medical AI.
• This approach enhances the transparency and trustworthiness of AI diagnostics.
• The method provides a robust framework for generating precise visual explanations in medical imaging.