KARNet: A Novel Deep-Learning Approach for Dementia Stage Detection in MRI Images

Introduction Accurate detection and staging of dementia are crucial for early intervention and effective patient management. Magnetic resonance imaging (MRI) serves as a valuable diagnostic tool, and deep learning models have the potential to enhance its accuracy and efficiency. Objective This study introduces KARNet, a novel deep-learning framework that integrates the Kolmogorov-Arnold network (KAN) architecture with a modified residual neural network (ResNet-18) and principal component analysis (PCA) to classify four stages of dementia: non-demented, very mild dementia, mild dementia, and moderate dementia. Methods To optimize model performance, we employ transfer learning by modifying a pre-trained ResNet-18 as a feature extractor, followed by a KAN layer as the classifier. PCA is adopted to reduce training time and computational complexity. Additionally, an ablation study and hyperparameter optimization are conducted to evaluate the robustness of the proposed model and improve performance. Results Experimental results demonstrate that KARNet achieves a classification accuracy of 98.5%, outperforming the existing state-of-the-art models. Evaluation on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset confirms its effectiveness in enhancing classification accuracy and model reliability for dementia staging. Conclusion The findings suggest that KARNet is a promising deep-learning framework for the early diagnosis and monitoring of dementia stages using MRI, offering a potential advancement in automated dementia assessment.