Development of a diagnostic aid system for neurodegenerative pathologies: Application to the detection of Alzheimer’s disease

Abstract

Alzheimer’s disease (AD) is a degenerative disorder and one of the most widespread forms of dementia, with no current cure available. This absence of a cure has led the medical field to focus on managing the symptoms of the disease. However, its progressive nature complicates the identification of disease stages, often requiring years of expertise. Consequently, computer-aided diagnostic systems are essential to assist clinicians in accurately defining disease stages, enabling more targeted and effective treatments. Given the lack of a cure, early detection of AD, particularly at the Mild Cognitive Impairment (MCI) stage, is crucial to slow or halt disease progression. However, distinguishing MCI symptoms from normal aging remains challenging, even with MRI imaging, due to the subtle differences across MCI substages. This thesis focuses on accurately classifying AD stages, emphasizing early-stage detection. Two MRI databases were utilized, leading to four methodologies addressing specific challenges while fulfilling the research objectives. The first system combines the Histogram of Oriented Gradients (HOG) with the Bias Correction Fuzzy C Means algorithm and machine learning classifiers, achieving an accuracy of 96.8% for the first database and 96% for the second. The second system shifts to the frequency domain, employing the Fast Finite Shearlet Transform (FFST) and Gray Level Co-occurrence Matrix (GLCM) angles, resulting in 72% accuracy for the first database. Building on these approaches, the third system leverages inductive transfer learning with layer-wise fine-tuning of ten pre-trained models. The best results were achieved using the Xception architecture, yielding 85.19% accuracy for the first database and 77.23% for the second using VGG19. Finally, the fourth methodology integrates machine learning and deep learning by automatically extracting features and refining them using Bayesian optimization. It achieves an accuracy of 98.45% for the first database and 78.54% for the second. These methodologies address critical research questions and highlight the importance of feature quality, hyperparameter optimization, and data augmentation techniques in medical imaging. The findings underscore the potential of advanced computer-aided diagnostic systems to enhance the detection and staging of Alzheimer’s disease.