LIGHT WEIGHT DEEP LEARNING FRAMEWORK FOR DEMENTIA STAGE IDENTIFICATION USING CORONAL MRI SCANS

Abstract

Dementia is a rapidly increasing global health problem and refers to a range of progressive neurodegenerative conditions that result in the loss of memory, cognition and independent functioning. Dementia currently affects more than 55 million people globally and is expected to affect 139 million more by 2050, making the need for efficient, accurate and scalable diagnostic techniques more urgent than ever. Another important problem is the extreme class imbalance found in typical datasets for dementia imaging, especially the clinical stage Moderate-Demented (MD), which leads to a class-imbalanced diagnostic behavior of models trained on the conventional way. In this thesis, LW-DMNet (LightWeight Dementia MRI Network) is introduced, a parameter-efficient deep learning network designed for automated multi-class demen tia stage classification using coronal MRI images, which is clinically feasible. LW DMNet stands for a lightweight custom Convolutional Neural Network (CNN) archi tecture designed from first principles for the features extracted using coronal brain MRI, but using no pre-trained weights. It is built using four hierarchical convolutional blocks, a global average pooling layer, and a three-layer fully connected classification head, with only 13.39 million trainable parameters, which is about 90% fewer than the number of parameters in VGG16 (approximately 138 million). In order to solve the class imbalance problem in a principled way, the hybrid re sampling technique SMOTETomek is used. It combines the Synthetic Minority Over sampling Technique (SMOTE)tocreatenewsyntheticsamplesfortheunder-represented classes by interpolating them in feature space, with the Tomek Link removal tech nique to remove samples in ambiguous regions at the class boundary to produce a balanced and well-separated training distribution. The experiment results validate LW DMNet with an accuracy of 98.44% and F1-scores of 1.00 for the critically important Moderate-Demented class, and 98.19% for macro-averaged scores, despite a small number of 64 training instances for this class. The discriminative power of the pro posed model is further validated by the macro-average AUC, which is 0.998. These findings motivate LW-DMNet as a clinically feasible, computationally efficient ap proach to dementia staging based on brain images.