Reducing annotation burden in medical imaging with ADGNET: A semi-supervised deep learning strategy

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Alzheimer's disease (AD) diagnosis relies heavily on medical imaging, but limited annotated data poses a challenge for deep learning models.
• Existing methods often struggle with class imbalance and effective feature extraction from sparse datasets.

Purpose of the Study:

• To introduce ADGNET, a semi-supervised framework for improved Alzheimer's disease diagnosis using magnetic resonance imaging (MRI).
• To leverage a dual-task learning approach for joint image reconstruction and classification, enhancing feature representation with limited annotations.

Main Methods:

• ADGNET integrates a residual backbone with attention, an encoder-decoder for unsupervised learning, and a classification branch with focal loss.
• The framework optimizes shared feature representations for both image reconstruction and AD classification tasks.
• Utilized two public MRI datasets: KACD (2D) and ROAD (3D) for model validation.

Main Results:

• ADGNET demonstrated significant performance improvements over state-of-the-art methods (ResNeXt WSL, SimCLR) by 4.1% (KACD) and 7.2% (ROAD) across six metrics.
• Interpretability analyses (Grad-CAM, attention visualization) confirmed the model's focus on clinically relevant regions like the hippocampus and temporal lobes.
• The model showed strong correlation (r=0.67, p<0.001) between its learned features and established Alzheimer's pathology.

Conclusions:

• ADGNET offers an efficient and effective solution for few-shot medical image analysis in Alzheimer's disease diagnosis.
• The framework exhibits strong generalization capabilities across multi-modal medical imaging data.
• The joint optimization of reconstruction and classification enhances feature learning and diagnostic accuracy.