Recently, a considerable number of investigations on deep learning methodologies for early identification of Alzheimer's disease (AD) have been reported in the literature. For accurate diagnosis, the currently used feature modelling algorithms must be improved. Such the gap is addressed by proposing a dual stream of dilated convolutional neural networks. The architecture uses Canny edge maps and MRI scans as two samples’ cues to extract fine-grained spatial characteristics. To enhance the quality of feature modelling process, the high-level features from the dual streams are combined. To predict illness stages, the fused features are flattened and tightly connected with 256 neurons. These neurons are then fully connected to an output layer for AD diagnosis. For testing and validation, the publicly accessible AD dataset is considered in this article to verify the performance of the proposed architecture. With a 95.93% accuracy rate, the suggested model outperforms the current best practices. Additionally, an ablation study on the architecture is also included to illustrate how each stream behaves for each sample cue.