Classification of Alzehimer’s Disease from 2D Brain MRI Slices Using Convolutional Neural Networks

Abstract

The diagnosis of Alzheimer's disease and related dementias relies on both clinical assessment and brain imaging, but it is unclear which data type provides a more reliable predictive signal for automated classification. This study compares clinical/volumetric features with MRI images for classifying OASIS-2 subjects as Demented or Nondemented, using 150 subjects (373 visits) with subject-grouped cross-validation to prevent a patient's repeat visits from spanning both training and test sets. Clinical/volumetric features (age, sex, education, socioeconomic status, Mini-Mental State Examination (MMSE) score, FreeSurfer brain volume measures, estimated Total Intracranial Volumne (eTIV), Atlas Scaling Facotor (ASF), and normalized whole-brain volume (nWBV)) were evaluated via logistic regression, while raw MRI slices (136 labeled images) were evaluated using both a CNN and a VGG-16 architecture, both trained from scratch with tuned optimizer hyperparameters. Logistic regression reached 70% accuracy (69% Demented recall), with nWBV as the strongest individual predictor; accuracy rose to 84% (81% recall) when the MMSE cognitive test score was included. MRI-based classifiers performed substantially worse: the CNN reached 58% accuracy under 5-fold cross-validation, only marginally above a 53% majority-class baseline, and the best-tuned VGG-16 configuration reached 57%. While these results indicate that clinical/volumetric features produced better classification performance in this study than MRI images, CNN training on one MRI slice per patient likely reduced the performance of these models. As demonstrated through Grad-CAM analysis, CNNs can provide insight into the specific brain regions contributing to Demented predictions and aid in identifying structural patterns associated with dementia.