Vision Transformers (ViTs) have shown strong potential in medical image classification due to their ability to model long-range dependencies. Despite this advantage, the deterministic nature of the standard Multi-Head Attention (MHA) mechanism can lead to overfit ting and reduced robustness, especially when working with noisy and heterogeneous medical datasets. To address this issue, we introduce a modified attention mechanism called Noise-Injected Multi-Head Atten tion (NIMHA), which integrates controlled Gaussian noise into the key and value projections of MHA. This stochastic regularization approach enhances feature learning and model generalization while maintaining computational efficiency and compatibility with existing ViT architec tures. We evaluate NIMHA on two public datasets: Brain Tumor MRI and CT Kidney. Experimental results show that ViTs with NIMHA out perform baseline ViTs in classification accuracy, particularly on the more complex Brain Tumor MRI dataset. In addition, models with NIMHA exhibit more stable training behavior and faster convergence. Attention map analysis further reveals that the proposed method promotes a more distributed focus, improving the model’s ability to generalize to diverse clinical data. These findings suggest that incorporating noise-based reg ularization into attention mechanisms is a practical strategy to enhance the robustness and reliability of ViT-based models for medical imaging tasks.