Abstract
Alzheimer's disease is a chronic clinical condition that is predominantly seen in age groups above 60 years. The early detection of the disease through image classification aids in effective diagnosis and suitable treatment. The magnetic resonance imaging (MRI) data on Alzheimer's disease have been collected from Kaggle which is a freely available data source. These datasets are divided into training and validation sets. The present study focuses on training MRI datasets using TinyNet architecture that suits small-scale image classification problems by overcoming the disadvantages of large convolutional neural networks. The architecture is designed such that convergence time is reduced and overall generalization is improved. Though the number of parameters used in this architecture is lesser than the existing networks, still this network can provide better results. Training MRI datasets achieved an accuracy of 98% with the method used with a 2% error rate and 80% for the validation MRI datasets with a 20% error rate. Furthermore, to validate the model-supporting data collected from Kaggle and other open-source platforms, a comparative analysis is performed to substantiate TinyNet's applicability and is projected in the discussion section. Transfer learning techniques are employed to infer the differences and to improve the model's efficiency. Furthermore, experiments are included for fine-tuning attempts at the TinyNet architecture to assess how the nuances in convolutional neural networks have an impact on its performance.