The complex patterns of the neuroimaging data are analyzed successfully with bio-medical imaging applications. The patients with/without AD can be discriminated effectively through several biomedical imaging modalities such as, sMRI, fMRI, PET and so on. In this paper, brain images from structural MRI (sMRI) are used for better categorization of 3 subjects namely, NC (Normal Control), MCI (Mild Cognitive Impairment) and AD (Alzheimer’s disease). Moreover, ambiguous training data employed for the discrimination of subjects may mislead the classifier to take incorrect decisions and in turn degrades the classification performance. In order to recover these hurdles, we propose an automated reliable system for the detection of AD affected patients accurately from the brain images of sMRI. The proposed system is a multi-stage system comprising four key phases namely, i) pre-processing, ii) feature extraction, iii) feature selection and iv) detection phase. In the initial phase, ROI regions related to Hippocampus (HC) and Posterior Cingulate Cortex (PCC) from the brain images are extracted using Automated Anatomical Labeling (AAL) method. In the feature extraction stage, important texture and shape features are extracted from HC and PCC involved in three brain planes. Nearly, 19 highly relevant AD related features are selected through a multiple-criterion feature selection method. It should be noted that, the class labels when explored manually consumes more time and turns to be an expensive process. Therefore, it is essential to construct an automatic method to identify irrelevant samples in the training data to enhance the decision-making process. With this in mind, at the detection phase, a novel classification technique is proposed by combining the Kernel fuzzy c-means clustering (i.e. unsupervised learning technique) and Back-propagation artificial neural network (i.e. supervised learning technique) to categorize NC, MCI and AD from the brain images of sMRI. This proposed KFCM based BPANN algorithm can improve the classification performance by removing the suspicious training samples. The proposed frameworks efficiency is evaluated with the ADNI subset and then to the Bordeaux-3 city dataset. The experimental validation of our proposed approach attains an accuracy of 97.63%, 95.4%, 96.4% for the most challenging classification tasks AD vs NC, MCI vs NC and AD vs MCI, respectively.