This work explores diverse architectures for deep convolutional neural networks (CNNs). It analyzes the influence of domain-specific transfer learning on distinguishing between patients with Alzheimer’s Disease (AD), Mild Cognitive Impairment (MCI), and cognitively normal (CN) subjects. Training relied on 1118 3D FDG-PET brain scans acquired through the ADNI initiative. We developed standalone binary classifiers, a multiclass model, and two different chains composed of binary classifiers for sequential multiclass decision-making. Chain A started by separating cognitively normal (CN) individuals from all non-CN cases, then distinguished Alzheimer’s disease (AD) from mild cognitive impairment (MCI). Chain B reversed the initial step, first identifying AD against non-AD subjects before classifying MCI versus CN. Every classifier variant was trained from scratch and with domain-specific transfer learning initialized from Med3D weights. All configurations matched or approached current state-of-the-art levels. Chain A stood out, recording 96% accuracy, 95% F1 score, and 99% AUROC. In contrast, domain-specific transfer learning often degraded performance across most models, reducing accuracy by up to 55%. Overall, the study underscores the value of binary classifier chains while prompting reconsideration of domain-specific transfer learning practices.
