ADLi-Net: A Learnable Dilation Deep Learning Framework for Accurate Alzheimer’s Disease Stage Classification from Brain Magnetic Resonance Images

Alzheimer’s Disease (AD) is a leading cause of dementia because of its advanced neurodegenerative nature. However, conventional Deep Learning (DL) models that use fixed dilation are struggle to capture both fine-grained and global structural changes, thereby leading to suboptimal feature representations. In this research, a novel framework named Alzheimer’s Disease Learnable Dilation Network (ADLi-Net) is proposed, which integrates DenseNet-201 with a Learnable Atrous Spatial Pyramid Pooling (ASPP) module for multi-class AD classification using Magnetic Resonance Imaging (MRI). The learnable ASPP dynamically learns through backpropagation during training, thereby providing a model for extracting context-aware features from complex brain structures. Preprocessing techniques, such as Contrast Limited Adaptive Histogram Equalization (CLAHE) and min-max normalization are utilized to enhance image contrast and standardize pixel intensity values thereby ensuring consistent input for further process. The proposed ADLi-Net is estimated on the ADNI dataset containing over 23,000 T2-weighted MRI scans categorized into six cognitive states. The experimental results show that ADLi-Net achieves better performance, with an overall accuracy of 98.19% on ADNI dataset. The comparative analysis and ablation study ensured the efficiency of learnable dilation rates, weighted loss functions, and optimal validation strategies. Therefore, this research demonstrates that dynamically tuned multi-scale spatial filtering enhances AD classification.