SMR-Net: A Multi-Resolution Neuroanatomical Representation Learning Framework for Early Diagnosis of Alzheimer's Disease

SMR-Net is a multi-resolution neuroanatomical representation learning framework for the early diagnosis of Alzheimer's Disease (AD). It achieves high-precision modeling of AD-related brain structural abnormalities through cross-resolution fusion, slice-level attention aggregation, and graph neural network reasoning. This framework aims to address the problems of strong subjectivity in traditional diagnosis and insufficient single-scale analysis, providing a new tool for early AD diagnosis.

Alzheimer's Disease is the leading cause of dementia among the elderly worldwide. Early diagnosis is crucial for delaying disease progression, but traditional methods rely on empirical judgment, which is highly subjective and struggles to capture subtle structural changes. Magnetic Resonance Imaging (MRI) provides high-resolution brain information, but effectively extracting multi-scale neuroanatomical features is a key challenge in computational neuroscience—single-scale analysis cannot fully capture the complexity of lesions.

The core innovations of SMR-Net include: 1. Cross-resolution feature fusion: Decompose MRI into multi-scale feature maps—low resolution captures macrostructures (e.g., ventricular enlargement), while high resolution focuses on local textures (e.g., hippocampal blurring). Adaptive weights dynamically adjust the contribution of each feature. 2. Slice-level attention aggregation: Calculate attention weights along axial, coronal, and sagittal planes, then aggregate key slice information with weights, focusing on brain regions affected early in AD such as the hippocampus. 3. Graph neural network reasoning: Model the brain as a graph structure (nodes are brain regions, edges are connections), and learn topological relationships between brain regions and pathological propagation patterns via graph convolution.

Technical innovations of SMR-Net: 1. Multi-resolution architecture avoids the limitations of single-scale analysis. 2. Slice-level attention provides an interpretable mechanism (doctors can understand decision-making basis via heatmaps). 3. Graph neural networks explicitly model inter-brain region relationships, compensating for the lack of global correlation in convolutional networks. 4. End-to-end training eliminates the need for manual feature engineering, lowering the threshold for use.

Application value of SMR-Net: 1. Clinical aspects: Assist radiologists in accurate image interpretation, improving the consistency and accuracy of AD diagnosis; identify the risk of AD conversion in patients with Mild Cognitive Impairment (MCI) to enable early intervention. 2. Research aspects: Explore AD neuroanatomical mechanisms through multi-resolution analysis and graph reasoning, and discover new imaging biomarkers. 3. Versatility: After adjustment, it can be applied to image analysis of other neurodegenerative diseases such as Parkinson's Disease and Frontotemporal Dementia.

SMR-Net demonstrates strong performance in AD modeling tasks through the combination of cross-resolution fusion, attention mechanisms, and graph neural networks. With the expansion of data scale and algorithm optimization, such AI-assisted diagnostic tools are expected to become powerful assistants for neurologists, ultimately benefiting a large number of patients.