Early detection of Alzheimer’s disease (AD) is important for timely intervention and slowing disease progression. As Magnetic Resonance Imaging (MRI) provides structural brain information used by medical specialists in AD assessment, researchers have increasingly applied deep learning models such as ResNet to MRI data to support automated AD detection. However, most existing studies focus primarily on modifying network architectures, with limited attention given to the impact of varying design and training parameters on model performance. This study proposes an ADResNet framework that combines ResNet feature extraction with a Bidirectional Gated Recurrent Unit (Bi-GRU) for MRI brain scans. Three ResNet variants, namely ResNet-18, ResNet-34, and ResNet-50, were developed and systematically evaluated four pooling techniques: Global Max Pooling, Global Average Pooling, Concatenation Pooling, and Attention Pooling. The experiments further examined the impact of varying batch sizes and learning rate strategies under consistent experimental conditions. The results showed that AP consistently achieved the best overall performance across the evaluated models. Increasing the batch size improved model generalisation by reducing the training–validation gap, while the variable learning rate strategy produced more stable convergence than the fixed learning rate approach. The best overall performance in terms of accuracy, training stability and convergence was achieved with the ResNet-50 model using Attention Pooling, a batch size of 32, and variable learning rate scheduling. The findings further demonstrate that design and training parameters can significantly impact the performance and stability of ResNet-based models for AD detection.