Julia Westermark will present her master's thesis.

Title: Multi-modal Alzheimer’s Prediction with Explainable AI

Author: Julia Westermark

Supervisors: Anders Heyden and Sanna Persson (KTH)

Examiner: Niels Christian Overgaard

Abstract:

Although deep learning models achieve high performance in Alzheimer’s disease classification, the lack of transparency in black box models limits their clinical acceptance, highlighting a critical need for explainable AI. In this study, an intrinsically interpretable prototype based network was implemented for both uni-modal and multi-modal imaging techniques. Using magnetic resonance imaging (MRI) and positron emission tomography (PET) datasets, the model classified subjects into Alzheimer’s disease, mild cognitive impairment, and cognitively normal. The model was evaluated using both performance and explainability metrics to analyze the model’s decisionmaking process. The uni-modal MRI model showed a balanced accuracy of 58.1% and rarely predicted mild cognitive impairment. The explanations did not show compactness but the prototypes showed some clinical relevance. On the other hand, the PET model demonstrated a higher balanced accuracy of 70.3% and better explainability, but the model was still lacking in its explainability properties. The multi-modal approach achieved a lower performance (51.5%), as the network failed to utilize PET prototypes, essentially treating the limited PET data as noise. While the PET model yielded better performance than the MRI model, the multi-modal model failed to leverage this advantage due to modality imbalance.