Ανάπτυξη μοντέλων μηχανικής μάθησης για την ανίχνευση νευροεκφυλιστικών παθήσεων σε νευροαπεικονιστικά δεδομένα

Abstract

Detecting brain disorders is critical to understanding and managing changes that affect human consciousness and behavior. Normative modeling offers a breakthrough approach by enabling the detection of pathologies through comparison with normal brain functional patterns. The integration of artificial intelligence and machine learning, in particular autoencoders, greatly enhances the accuracy and efficiency of this technique. This thesis presents the development of two normative modeling frameworks: one using an autoencoder architecture and the other using a variational autoencoder. These models were trained on data from healthy individuals in the UK Biobank. After adjustment for confounders using linear regression and standard scaling, a normative pattern was established. The normative pattern was based on healthy subjects. The models were then tested on semi-synthetic data with simulated atrophy to assess deviations using reconstruction error. Furthermore, the models were evaluated on the ADNI dataset, which includes individuals with normal cognitive function, mild cognitive impairment, and Alzheimer’s disease. The results showed that these models can effectively discriminate between healthy and pathological brain states, paving the way for early diagnosis and treatment of brain disorders. In addition, the models can classify diseases with minimal labeled data for training, providing significant advantages over traditional methods that require extensive labeled datasets.