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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate dementia staging is vital for timely intervention and patient care.
• Magnetic resonance imaging (MRI) is a key diagnostic tool for dementia.
• Deep learning offers potential to improve MRI-based dementia detection accuracy and efficiency.

Purpose of the Study:

• Introduce KARNet, a novel deep learning framework for classifying four dementia stages (non-demented, very mild, mild, moderate).
• Integrate Kolmogorov-Arnold Network (KAN) with ResNet-18 and Principal Component Analysis (PCA) for enhanced dementia classification.
• Evaluate KARNet's performance against state-of-the-art models using the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.

Main Methods:

• Utilize transfer learning with a pre-trained ResNet-18 as a feature extractor.
• Employ a KAN layer as the classifier for dementia staging.
• Apply PCA for reducing computational complexity and training time, alongside ablation studies and hyperparameter optimization.

Main Results:

• KARNet achieved a high classification accuracy of 98.5%.
• The model demonstrated superior performance compared to existing state-of-the-art methods.
• Effectiveness was confirmed on the ADNI dataset, enhancing classification accuracy and reliability.

Conclusions:

• KARNet presents a promising deep learning approach for automated dementia staging using MRI.
• The framework facilitates early diagnosis and monitoring of dementia progression.
• This represents a potential advancement in automated dementia assessment tools.