On February 13th, the student Anastasiia Mikhailova, from the PhD program in Electrical and Computer Engineering, will defend her thesis titled “Computational Model of Visual Memory for Early Diagnosis of Dementia,” supervised by Professor José Santos-Victor. The defense will take place in the PA-3 Amphitheater (Floor -1 of the Mathematics Pavilion) at IST, at 2:30 PM.

The jury will be chaired by Professor Pedro Lima and composed of José Santos-Victor, from the Department of Electrical and Computer Engineering at Técnico, João Sanches and Paulo Correia, from Técnico, Nuno Garcia, from the Faculty of Sciences at the University of Lisbon, Luís Alexandre, from the Faculty of Engineering of Beira Interior, and Karla K. Evens from the University of York, UK.

Abstract:

This thesis presents the exploration of long-term visual memory mechanism in young adults and Mild Cognitive Impairment (MCI) along with its modelling to enhance early diagnosis of cognitive decline. First, it explores the correspondence between features extracted from early and late convolutional neural network (CNN) layers with saliency and visual memory schemas. Second, it examines long-term memory for scenes in young people by analysing the impact of image features to understand the mechanism of image features' influence on long-term visual memory. Third, it extends this analysis to MCI and elderly people to further understand the mechanism of MCI memory impairments. Fourth, it models the image diagnosticity for MCI by using the image memorability model. Finally, the thesis develops a diagnostic model for MCI based on gaze exploration patterns and image information by using CNNs to discriminate between MCI and healthy controls. The results reveal that features extracted by CNNs can approximate visual memory processes and present new findings on how specific image characteristics influence memory performance in healthy individuals and those with MCI. The empirical studies reveal different patterns of long-term visual memory in MCI or healthy populations, particularly in scene recognition and image feature processing. Additionally, the diagnostic model developed in this thesis shows promise in enhancing the accuracy and accessibility of MCI screening by integrating gaze and image data, which is
complemented by another model for the selection of diagnostic images.

This research contributes to cognitive psychology, computer science and engineering by extending our
understanding of visual memory and its impairment in MCI and proposing novel computational tools for early diagnosis.