Dementia is a common term used to say about memory loss and other mental-related problems—physical changes in the human brain cause it. There exist different types of dementia, and the commonly seen are Alzheimer’s and Parkinson’s diseases. For the past thirty years, many studies have been conducted on diagnosing dementia, its treatment, and monitoring assessments. The study's main aim is to design highly accurate systems to distinguish between different types of dementia. Many studies and classification models have been developed to diagnose significant types of dementia like Alzheimer’s and Parkinson’s diseases. But, for the other types of dementia, not many studies have been conducted because of the lack of availability of the dataset. So, the proposed system designed a generative deep convolution neural network (CNN) to diagnose different types of dementia with the same algorithm. This paper proposes a GAN (generative adversarial network) technology to generate virtual scan images for the other type of dementia except AD and PD to solve the problems caused by uniform distribution among the dataset categories. The input for the model must be balanced so that the patient data from all types of dementias are equally weighted. This also supports and emboldens the need for a larger dataset, enabling us to use synthetically generated data points. Synthetic data can be generated by making a model from the existing data, predicting new data points, and finally making sure they fit into the existing data distribution. Another viable strategy is to generate random data and tune the model until it can generate acceptable output, which falls into the required distribution. The dataset selected is a set of FDG-PET (fluorodeoxyglucose—positron emission tomography) scanned images with the label AD, PD, OD (Other Dementia), and HB (Healthy Brain) and a Deep CNN do the classification. This work used 1200 images in a PET image dataset, including AD, PD, and Healthy Brain images. The two distinct categories for training and testing of the total number of images are in the ratio of 7:3. The classification architecture of the proposed system achieves an overall accuracy of 97.7%.