Abstract
Background/Objectives: Alzheimer's disease (AD) is a progressive neurodegenerative disorder causing memory, cognitive, and behavioral decline. Early and accurate diagnosis is critical for timely treatment and management. This study proposes a novel hybrid deep learning framework, GLCM + VGG16 + FMO + CNN-LSTM, to improve AD diagnosis using MRI data. Methods: MRI images were preprocessed through normalization and noise reduction. Feature extraction combined texture features from the Gray-Level Co-occurrence Matrix (GLCM) and spatial features extracted from a pretrained VGG-16 network. Fisher Mantis Optimization (FMO) was employed for optimal feature selection. The selected features were classified using a CNN-LSTM model, capturing both spatial and temporal patterns. The MLP-LSTM model was included only for benchmarking purposes. The framework was evaluated on The ADNI and MIRIAD datasets. Results: The proposed method achieved 98.63% accuracy, 98.69% sensitivity, 98.66% precision, and 98.67% F1-score, outperforming CNN + SVM and 3D-CNN + BiLSTM by 2.4-3.5%. Comparative analysis confirmed FMO's superiority over other metaheuristics, such as PSO, ACO, GWO, and BFO. Sensitivity analysis demonstrated robustness to hyperparameter changes. Conclusions: The results confirm the efficacy and stability of the GLCM + VGG16 + FMO + CNN-LSTM model for accurate and early AD diagnosis, supporting its potential clinical application.