Early detection of moderate cognitive impairment using magnetic resonance imaging (MRI) is critical for dementia therapy. Deep learning architecture produces good results in such studies. Algorithms need a huge number of annotated datasets to train a model. We avoid this obstacle in our study by employing layer-wise transfer learning and tissue segmentation of brain images to detect Alzheimer's disease in its early stages (AD). For layer-wise transfer learning, the VGG architecture family with pre-trained weights was employed. The proposed model distinguishes between normal control (NC), early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI), and Alzheimer's disease (AD). The Alzheimer's Disease Neuroimaging Initiative (ADNI) database was accessible by 85 patients with NC, 70 patients with EMCI, 70 patients with LMCI, and 75 patients with AD in this research. Each patient's grey matter (GM) tissue was removed utilising tissue segmentation. Preprocessing data are utilised to assess the proposed technique, which obtains the highest rates of classification accuracy on AD vs. NC (98.73%) and EMCI vs. LMCI patients (83.72%), while remaining classes accuracy is more than 80%. A comparison with earlier studies revealed that the proposed model beat the state-of-the-art models in terms of testing precision.