Explainable prototype-based image classification using adaptive feature extractors in medical images.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Prototype-based classifiers offer explainable AI (XAI) by using data samples (prototypes) for classification.
• Existing methods often use generic feature extractors, limiting prototype representativeness and classifier performance.
• This gap hinders effective application in specialized domains like medical imaging.

Purpose of the Study:

• To propose a novel cluster-oriented training strategy for enhanced prototype-based classifiers.
• To improve both the performance and explainability of AI models in image classification tasks.
• To address limitations in feature extraction for identifying representative prototypes.

Main Methods:

• Developed Explainable Prototype-based Image Classification (EPIC) with a novel Cluster Density Error (CDE) loss function.
• Fine-tuned feature extractors to preserve representative feature vectors in the latent space.
• Employed Principal Component Analysis (PCA) for feature vector dimensionality reduction.

Main Results:

• Achieved high classification accuracy (up to 95.01%) and Area Under the Curve (AUC) (0.992) on medical image datasets.
• Demonstrated superior explainability compared to existing prototype-based methods.
• Significantly reduced the number of prototypes required (by 98.38%) while improving performance.

Conclusions:

• The proposed EPIC method enhances prototype-based classifier performance and explainability.
• CDE loss and PCA integration effectively identify representative prototypes and reduce complexity.
• This approach shows promise for accurate and interpretable AI in medical image analysis.