A Comparative Analysis of Gradient-Based, Edge-Based, and Segmentation-Based Data Augmentation Methods for Early Diagnosis of Alzheimer’s Disease Using Neuroimaging Modalities and Deep Learning
Muhammad Dawood, Usman Rasheed, Waqas Ahmad, Ahsan Bin Tufail, Afnan Albahli
Alzheimer’s disease (AD) is a neurodegenerative disorder that causes progressive damage to brain neurons, leading to declines in cognitive and behavioral abilities. This deterioration often results in changes in personality and increasing difficulty in thinking and memory over time. Although there is no cure, early detection is crucial as it allows for more effective management and care. Advances in deep learning have significantly improved the accuracy of brain scan analysis for diagnostic purposes. In this study, we utilized the publicly available Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset consisting of subjects diagnosed with AD, Mild Cognitive Impairment (MCI), and Normal Control (NC). Each participant has either Magnetic Resonance Imaging (MRI) or Positron Emission Tomography (PET) neuroimaging data, ensuring representation across heterogeneous modalities. The research focuses on comparing gradient-based, edge-based, and segmentation-based data augmentation techniques for early AD detection using neuroimaging and deep learning approaches, particularly 3D Convolutional Neural Networks (3D CNNs). Various augmentation methods were applied, including directional gradient, azimuth gradient direction, numerical gradient, Sobel horizontal edge filter, superpixel oversegmentation, and Canny edge detection. These techniques are evaluated in both binary and multiclass classification tasks involving MRI and PET scans. The results indicate that optimal performance varied depending on the task and modality. For PET-based classification, directional gradient performed the best for AD vs. NC binary classification, achieving an accuracy of 87.24%, while Canny edge detection was most effective for AD vs. MCI binary classification and AD-MCI-NC multiclass classification tasks, achieving accuracies of 72.77% and 59.04%, respectively. For MCI vs. NC, the best result (accuracy = 64.32%) is achieved by combining azimuth gradient direction with Sobel filtering. In contrast, for the MRI-based AD vs. NC classification task, the highest performance is achieved without applying augmentation (balanced accuracy = 60.90%). This research confirms the efficacy of data augmentation methods in the early diagnosis of AD in clinical settings.