Triplet-Loss-Based Hybrid Siamese Convolutional Neural Network Model for Alzheimer's Disease Detection

Abstract

Alzheimer's disease (AD) is a neurological disorder that weakens the brain over time and affects memory and cognition. Due to the more comprehensive view of changes occurring in the brain, multimodal imaging methods have become more useful in the diagnosis of AD and in tracking the disease's course over time. Furthermore, the models that are currently in use do not produce good results for AD identification. Because of the intricate structure of the brain, these models face problems like overfitting, complicated modeling, and incorrect categorization that result in multi-model data. To provide a solution, an effective triplet-loss-based hybrid Siamese convolutional neural network model for the detection of AD is introduced. Skull stripping is first used to pre-process the neuroimaging data, and then, data augmentation techniques such as rescaling, rotation, horizontal flipping, and vertical flipping are employed to balance the dataset. Following pre-processing, the Integrated Swin-based improved Generative U-Net model (ISIGU) will be used to carry out the segmentation process in order to identify the affected section of the brain specifically. Using a Triplet-Loss-Based Hybrid Siamese Convolutional Neural Network Model (THSCNN), characteristics are retrieved from the segmented magnetic resonance imaging images and used to classify the phases of AD. The Enhanced Sine chaos Archimedes Optimization Algorithm (ESCAO) is used to refine the hyperparameters for improved outcomes and to optimize the loss that occurs in the classification model. The evaluation results of the model achieve an accuracy of 99.67% in CN detection, 99.74% in MCI, 99.63% in EMCI, 99.87% in LMCI, and 99.61% in AD detection.