Over the past few decades, Machine Learning (ML) models have demonstrated their ability to harness biomedical data to enhance the healthcare system. ML-based medical devices can serve both as support tools for patients and as aids for physicians in diagnosis, prognosis, and treatment planning. After briefly exploring various medical applications, this work turns to a central challenge in the field: the use of ML for Alzheimer's Disease (AD), the most common form of dementia, where timely diagnosis is crucial for slowing disease progression. This contribution presents a comprehensive investigation of ML techniques for AD detection using 3D brain MRI. The first part examines how modeling design choices—particularly data augmentation strategies and network complexity—affect the reliability and predictive per-formance of 3D Convolutional Neural Networks (CNNs). Using low-resolution 1.5T MRI scans from the ADNI dataset, fifteen models are trained by combining three affine-based augmentation strategies with five CNN architectures of in-creasing depth. The results show that these choices can lead to up to 10% variation in accuracy. Applying affine trans-formations separately consistently improves performance, and the relationship between model depth and accuracy fol-lows a concave trend, with intermediate-complexity architectures yielding the best results. The best model reaches 87% accuracy on internal testing. However, when evaluated on an external dataset acquired with 3T scanners and different protocols, performance decreases by 16 percentage points, highlighting the significant impact of acquisition variability. The second part addresses this challenge directly by exploring Transfer Learning (TL) as strategy to mitigate data scar-city and distribution shifts caused by evolving MRI acquisition technologies. Two scenarios are considered: (A) leverag-ing historical 1.5T MRI data to enhance models trained on limited 3T scans, and (B) adapting 2D CNNs pre-trained on ImageNet (ResNet18/50/101) for 3D MRI processing when historical data are unavailable. In scenario (A), TL signifi-cantly boosts the baseline model’s accuracy from 63% to 99%. In scenario (B), fine-tuning models pre-trained on natu-ral images increases the baseline’s accuracy by up to 12 percentage points, achieving an overall accuracy of 83%.

Machine Learning, Transfer Learning, Healthcare, Alzheimer’s Disease, Magnetic Resonance Imaging