ALZHEIMER’S DISEASE CLASSIFICATION USING DEEP LEARNING WITH SOFT-NMS AND TRANSFORMER INTEGRATION

Abstract

Alzheimer’s disease is a debilitating neurological disorder that is characterized by thinking problems and forgetfulness. This can only be done by an accurate and early diagnosis. Traditional methods of diagnostics are time-consuming and prone to human error and manual analysis of MRI is such a procedure. In this work, the automated deep learning strategy of AD stage classification is based on the Transformer modules and Soft Non-Maximum Suppression and 3D Convolutional Neural Networks. The system analyzes 3D MRI scans of ADNI and other datasets in Kaggle. Some examples of the preprocessing methods include skull stripping, intensity normalizing, histogram equalization and volumetric scaling. Generalization is enhanced by applying data augmentation methods i.e., rotation, flipping and injection of noise. The hybrid system also applies the Soft-NMS technique to correct the prediction by being able to capture contextual and spatial information on brain structures. With a cross-entropy loss and Adam optimization hyper-parameters, the model beats baseline CNN and Res-Net by recording an accuracy of 94.1 and AUC-ROC of 0.967 after training with cross-entropy loss and optimized with Adam. The results demonstrate that this approach could be used to diagnosis AD in the clinical practices in a reliable scale able and early manner.