Novel Deep Learning Techniques for Early Detection of Neurological Disorders

Early detection is critical in neurological disorders such as Alzheimer’s disease and headaches, offering a window for interventions that may delay progression and improve outcomes. However, traditional diagnostic approaches often fail to provide timely and accurate assessments due to the multifactorial nature of these conditions. Multimodal neuroimaging, particularly magnetic resonance imaging (MRI) and positron emission tomography (PET), provides complementary structural and functional information that can enhance diagnostic accuracy if leveraged effectively. This dissertation presents a three-step deep learning framework that addresses key challenges in medical imaging analysis. First, a robust brain age prediction model is introduced that reformulates age estimation as an ordinal classification task. This mitigates systematic bias in regression-based methods and better captures deviations from normative aging, critical for identifying early signs of neurological disorders. Building on this, the second chapter proposes a novel quantization technique to improve latent representation learning in medical imaging. It leverages Mahalanobisdistance-based soft assignment to multiple codebook vectors, addressing limitations of traditional vector quantization such as codebook collapse and poor reconstruction diversity. When integrated into downstream tasks such as brain age prediction, the proposed quantization method maintains performance in low-resource settings. The final chapter focuses on enhancing early prognosis through a multimodal generative approach. Specifically, an MRI-guided latent diffusion model is developed to superresolve low-resolution PET scans. By conditioning the diffusion process on highresolution anatomical MRI, the model improves spatial resolution and quantification of pathological biomarkers such as beta-amyloid. Collectively, these contributions form a clinically motivated, resource-efficient, and multimodal deep learning pipeline for advancing the early detection of neurological disorders.