In this work, we propose a multi-hop graph attention module (MGA) to mitigate CNN’s drawbacks when it comes to obtaining nonlocal connections necessary for brain age prediction. Graph attention and the Markov property are used by MGA to capture both direct and indirect connections via feature map translation to graphs and distance-based rating calculation. Human brain age is a widely used biomarker of ageing to find variations in healthy people. Patients’ ages may be inferred from their brain imaging since ageing causes certain particular changes in the human brain. One of the most widely used state-of-the-art models for estimating human brain ages via transfer learning is the DenseNet model, which was made feasible by advancements in CNN's capacity to do classification and regression from pictures. The Brain Age Estimation (BAE) job is recommended to employ 2D-CNN as 3D-CNN has a significant memory overhead for this operation. This research includes many tests to minimise calculations without sacrificing overall performance. Different optimisers—Adam, Adamax, and Adagrad—are utilised for each model to feed slices from the three brain planes into the DenseNet model. The information extraction from images (IXI) MRI data source provided the chosen dataset. The assessment of each model with varying input sets is done using the MAE evaluation measure.