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

Prototype-based classifiers are a category of Explainable Artificial Intelligence methods that use representative samples from the data, called prototypes, to classify new inputs based on a similarity criterion. However, these methods often rely on pre-trained Convolutional Neural Networks as feature extractors, which may not be adapted for the specific type of data being used, thus not suited for identifying the most representative prototypes. In this paper, we propose a method named Explainable Prototype-based Image Classification, a cluster-oriented training strategy that enhances the performance and explainability of prototype-based classifiers. Our method uses a novel loss function, called Cluster Density Error, to fine-tune the feature extractor and preserve the most representative feature vectors in the latent space. We also use Principal Component Analysis-based approach to reduce the dimensionality and complexity of the feature vectors. We conduct experiments on four medical image datasets and compare the results with those from different prototype-based classifiers and state-of-the-art non-explainable learning methods. The proposed method demonstrated superior explainable capabilities and comparable classification performance to the compared methods. Specifically, the proposed method achieved up to 95.01% accuracy and 0.992 AUC using only 43 prototypes. This translated to an improvement in accuracy and AUC score of 21.54% and 9.06%, respectively, and a substantial reduction in the number of prototypes by 98,38%.