Alzheimer's Disease Classification Using Convolutional Neural Network-Random Forest Hybrid Architecture

Abstract

Worldwide, healthcare systems bear a heavy financial burden due to the progressive neurological illness known as Alzheimer's disease (AD), which impacts around 55 million people. In order to enable proper clinical action at the appropriate moment, corrective identification of the disease's stages is a crucial task. This paper compares two deep learning-based frameworks for classifying Alzheimer's disease from MRI brain scans: (1) a standalone four-block Convolutional Neural Network (CNN) with batch normalization, global average pooling, and dropout regularization; and (2) a hybrid approach that combines a Random Forest (RF) ensemble for classification of its 256-dimensional embeddings with a CNN backbone as a learned feature extractor. A three-class MRI dataset of samples—mildly demented, moderately demented, and non-demented—is used to train and evaluate both models. The dataset is divided into train, validation, and test sets in a 70:15:15 ratio after being preprocessed into 128x128 pixels. There is a 70:15:15 split between the train, validation, and test sets, and the dataset is pre-processed to 128x128 pixels. It is then loaded in parallel. Among the measures used for evaluation are accuracy, weighted precision, weighted recall, weighted F1-score, per-class specificity, false negative rate, false discovery rate, false positive rate, and negative predictive value. Additionally, there is the Matthews Correlation Coefficient (MCC). Experimental results reveal that the CNN+RF hybrid outperforms the CNN and demonstrates the usefulness of combining deep feature representation with ensemble learning for neuroimaging diagnosis.