VGG16 tops CNN benchmark with 95.3% accuracy for Alzheimer's MRI detection

Researchers at SPIE Optics + Photonics 2026 presented a systematic benchmark of convolutional neural networks for Alzheimer's disease detection from single-view MRI scans. The study, led by Hiram Zuniga, Ulises Orozco-Rosas, and Kenia Picos, evaluated ten architectures across the ResNet, DenseNet, MobileNet, EfficientNet, and VGG families under a consistent held-out test split. They introduced a two-stage transfer learning and fine-tuning pipeline, training on a class-balanced subset of 3,900 images derived from the OASIS dataset, which contains 86,437 single-view MRI brain scans labeled into four Alzheimer's stages: Non-Demented, Very Mild Dementia, Mild Dementia, and Moderate Dementia.

VGG16 emerged as the top performer, achieving 0.9637 validation accuracy and 0.9533 test accuracy — a notable result given its simpler architecture compared to modern deep networks. The paper also highlighted a critical clinical challenge: all ten models consistently struggled to distinguish between the Non-Demented and Very Mild Dementia stages, the earliest detectable transition. This suggests that while CNNs are highly accurate for moderate-to-severe Alzheimer's detection, early-stage classification remains a bottleneck. The authors' open benchmarking protocol provides a standardized way for future research to compare model performance, potentially accelerating the development of reliable screening tools for early intervention.