Proto-Caps: interpretable medical image classification using prototype learning and privileged information

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Explainable AI (xAI) is crucial for high-risk applications like medicine.
• Understanding AI decision-making is essential for diagnostic and therapeutic support systems.
• Current xAI methods may lack intuitive explanations for clinical evaluation.

Purpose of the Study:

• To introduce Proto-Caps, an intrinsically explainable model for image classification.
• To provide intuitive and comprehensive explanations for AI-driven medical decisions.
• To enhance the trustworthiness and performance of AI in medical contexts.

Main Methods:

• Developed Proto-Caps, a novel intrinsically explainable model for image classification.
• Utilized human-defined visual prototypes to explain model decisions.
• Evaluated performance on two public medical image datasets.
• Assessed explanation truthfulness by analyzing prediction-explanation alignment.

Main Results:

• Proto-Caps demonstrated superior performance compared to existing explainable AI approaches.
• The model's explanations, based on visual prototypes, were found to be truthful and aligned with predictions.
• Optimal model settings were identified through extensive hyperparameter studies.
• Incorporating explainability did not compromise model performance.

Conclusions:

• Proto-Caps offers a promising approach to intrinsically explainable AI in medical image classification.
• The use of visual prototypes enhances understanding and trust in AI diagnostic tools.
• Combining xAI with high performance is achievable, paving the way for safer clinical AI adoption.