GGDA-net: geometry-guided deformable attention network for Alzheimer's disease image classification

Convolutional neural networks (CNNs) have achieved remarkable success in medical image analysis, including Alzheimer's disease (AD) classification. However, conventional convolution operations rely on fixed sampling patterns, and most existing attention mechanisms primarily focus on feature responses while neglecting spatial sampling geometry, limiting their ability to capture structural variations in brain images.

To address these limitations, this paper proposes a Geometry-Guided Deformable Attention Network (GGDA-Net) for medical image classification. The proposed framework integrates Linear Deformable Convolution (LDConv) with a Geometry-Aware (GA) Attention mechanism to jointly model feature semantics and spatial geometry. Specifically, LDConv introduces adaptive spatial sampling through learnable offsets, enabling flexible modeling of geometric deformations in brain structures, while the GA attention exploits the resulting geometric cues to guide the network toward more informative anatomical regions.

The experimental results show that the accuracy rates on the two datasets reached 99.38 and 99.16% respectively, which are superior to the existing most advanced algorithms. At the same time, the model maintains a compact size and has a relatively low computational complexity. These results highlight the effectiveness of feature learning based on geometric perception in medical image analysis and Alzheimer's disease diagnosis.