<p>Early recognition of Alzheimer’s disease (AD) and its precursor state, mild cognitive impairment (MCI), is pivotal in interrupting the progression of the disease and providing suitable treatment. AD is a gradual, irreversible neurological disorder that profoundly impacts memory and cognitive functioning. The recent development of deep learning techniques has enhanced the efficacy of AD recognition. However, challenges include insufficient training of DL models, numerous learnable parameters, and limited feature information of an image from a single-source DL architecture. To address these challenges, we proposed a new network-level fused DenseInc226 layered deep learning architecture for AD prediction from the MRI scans. The proposed architecture is based on two sub-architectures: customized 10-block IncDense self-attention (CMA-IDSA) and customized multi-level dense modules with self-attention (CMDM-SA). Both architectures are fused at the network level and obtain a new fused network known as DenseInc226. The data augmentation was performed at the initial step of the training data. The training data is employed to train the proposed fused architecture with manual hyperparameter selection. The trained model is employed in the testing phase, and features are extracted using a testing set from the depth concatenation layer. To classify extracted features, we design a single-layered shallow neural network that classifies the features into the relevant Alzheimer class. The experimental process has been conducted on two publicly available datasets, such as Alzheimer’s Cognitive 5-Class MRI (AC5C-MRI) and Alzheimer’s MRI 4-Class (AMRI-4C), and obtained improved accuracy of 93.9% and 99.7%, respectively. Detailed ablation studies and comparisons with the proposed fused architecture with recent techniques show the AD classification’s improved accuracy and precision rate.</p>