Alzheimer's Disease Prediction Algorithm Based on Group Convolution and a Joint Loss Function.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• 18F-fluorodeoxyglucose positron emission tomography (18F-FDG PET) is effective for Alzheimer's disease (AD) prediction.
• Current PET imaging faces challenges like indistinct features, low signal-to-noise ratios, and artefacts, impacting accuracy, especially for mild cognitive impairment (MCI).

Purpose of the Study:

• To develop an improved AD prediction algorithm using advanced deep learning techniques.
• To enhance the accuracy and reliability of AD prediction from PET images, particularly for MCI patients.

Main Methods:

• A group convolutional backbone network (ResNet18-based) was designed for multi-channel lesion feature extraction.
• A hybrid attention mechanism was incorporated to focus on relevant diagnostic regions and learn feature weights.
• A joint loss function, combining cross-entropy with regularization, was proposed to prevent overfitting and improve generalization.

Main Results:

• The proposed algorithm demonstrated improved lesion feature extraction and enhanced learning of disease-relevant regions.
• Experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset showed a 2.4% increase in prediction accuracy compared to existing methods.
• The algorithm effectively addressed issues of indistinct features and artefacts in PET images.

Conclusions:

• The developed algorithm significantly improves AD prediction accuracy, especially for MCI.
• The combination of group convolution, attention mechanisms, and a joint loss function offers a more robust approach to neuroimaging-based disease prediction.
• This method shows promise for earlier and more accurate diagnosis of Alzheimer's disease.