Deep neural networks (DNNs) have shown remarkable performance in medical imaging tasks, yet their black-box nature hinders adoption in safety-critical clinical settings. Feature attribution maps offer a means of interpretability by highlighting input regions that influence model decisions. However, these maps are often noisy and sensitive to model parameters and input perturbations, limiting their reliability. We propose a robust attribution framework that utilize Test-Time Augmentation (TTA) to enhance the stability of feature attribution maps. Our method improves attribution consistency by aggregating attribution maps generated from multiple perturbed inputs, incorporating both stochastic noise and geometric transformations to capture a broader range of input variations. We experimentally demonstrate that the proposed method produces significantly more reliable results, with a \(35\%\) improvement on the Pointing Game metric, a standard for assessing attribution accuracy.