• 18F-FDG PET
• Alzheimer’s disease
• deep learning
• explainable artificial intelligence
• TCAV

In this thesis, we investigated the explainability of a convolutional neural network (CNN) for multiclass classification of volumetric 18F-FDG PET brain images from patients with Alzheimer's disease.
To achieve this, we applied a post hoc explanation method from the field of Explainable Artificial Intelligence: Testing with Concept Activation Vectors (TCAV). This concept-based approach is designed for the combined analysis of clinical data and biomedical images and it aims to evaluate how sensitive the model's predictions are to user-defined concepts when predicting each stage of Alzheimer's disease across all layers of the network.
The concepts of interest, identified by nuclear medicine specialists, focused on glucose hypometabolism in specific brain regions, where decreased glucose uptake is recognized as an indicator of Alzheimer’s disease.