Frontotemporal dementia (FTD) is a group of neurodegenerative disorders characterized by progressive damage to the frontal and temporal lobes of the brain. FTD includes notable changes in social behavior, personality, and language function difficulties. The primary FTD spectrum disorders include behavioral variant FTD, non-fluent variant primary progressive aphasia, and semantic variant. With an aging population, FTD prevalence increases, emphasizing the critical need for research, care, and awareness. This study introduces novel hierarchical classification methods that enhance model prediction capabilities. The dataset used in this work consists of neuropsychiatric tabular data from the Frontotemporal Lobar Degeneration Neuroimaging Initiative database. Innovative methods are needed to effectively leverage deep learning for FTD classification, especially when limited to tabular data. To address this challenge, we proposed an innovative technique, the Tensorised Image Generator (TIG), which transforms tabular data into grid-based image representations or tensors. The TIG algorithm optimizes tabular data visualization by ranking pairwise feature distances and then creating a distance matrix and strategically positioning features in the image to preserve spatial correlation among features. The algorithm further enhances the grid structure with line drawing and intensity adjustments. The results on these images demonstrate high accuracy in detecting FTD subtypes using a Convolutional Neural Networks ensemble with multi-layer perceptron, achieving a test accuracy of 88.89%, which is a 2% increment of hierarchical approach over flat machine learning methods and around 5% increment for combined hierarchical classification with image-based technique compared to best flat method across most metrics (accuracy, precision, recall, and F1-score).