ConvMix: An Improved Anatomical Features Based Model for Alzheimer's Disease Detection

Abstract

Progressive anatomical variation in the brain MRI plays an important role in the early detection of Alzheimer's disease. These structural changes modify anatomical characteristics and provide discriminative information. However, detection of multi-stage Alzheimer's disease remains challenging due to subtle inter-class variations and class imbalance in datasets. To address these issues, pre-processing techniques are employed to improve image quality and consistency. Furthermore, it is necessary to balance the Alzheimer’s dataset using borderline-SMOTE. The proposed model uses the convolutional network as the backbone feature extractor. It learns a hierarchical representation of anatomical features from brain MRI images by progressively extracting information through multiple stages of successive downsampling. By increasing its receptive field, it captures both local and global contextual information. The motivation for combining this Cross-Scale Feature Fusion is to further enhance these representations by integrating feature maps from multiple stages, thereby preserving fine grained anatomical information while incorporating global contextual features. This multi-level feature fusion improves the discriminative capability of the proposed model, resulting in multi-stage Alzheimer's disease detection. This proposed approach effectively detects the four stages of Alzheimer's disease while providing interpretable attention-region visualizations. The results obtained by this model achieve an accuracy of 98.75%.