Enhanced Local Feature Extraction for Alzheimer's Diagnosis using Res-VMamba Model

Alzheimer's Disease (AD) is a progressive neurodegenerative disorder that significantly affects memory and cognitive function. Early and accurate diagnosis plays a crucial role in delaying disease progression. Structural MRI imaging is a widely used modality for AD diagnosis due to its non-invasive and high-resolution properties. Although visual backbone architectures such as Visual Transformer and State Space Model (SSM) perform well in image classification tasks, current models still have shortcomings in local feature extraction. To address this challenge, this paper proposes a Res-VMamba model, which is an enhanced SSM architecture that integrates a residual convolutional patch embedding module inspired by the ResNet stem. This module improves local feature extraction by combining multi-scale convolutions, residual connections, and dual max pooling operations. The research evaluates this model on a multiclass MRI-based AD dataset containing three categories: 'Non Demented', 'Very Mild Demented', and 'Mild Demented'. Experimental results demonstrate that Res-VMamba outperforms state-of-the-art CNNs, Transformers, and the original VMamba model in classification accuracy (86.90 %) and F1-score (86.76), confirming the effectiveness of the proposed patch embedding design for fine-grained AD staging tasks. The study provides a more refined pathological identification tool for the diagnosis of $A D$ and has important practical application value.