The deep learning models achieved state-of-the-art results for various visual recognition tasks, and optimising these models using limited data is an active area of research. The use of rotationally transformed observations to augment the training data is a commonly used optimisation technique during training. In this work, we explore the use of rotationally transformed observations for evaluating the trained deep learning model. This work examines various angles of rotation to determine the optimal angle of rotation among them based on the confidence of prediction. The proposed approach is utilised to design an effective deep learning model for detecting age-related macular degeneration using fundus images. The experimental study on the AMDNet23 dataset suggests that the proposed approach’s performance is comparable to that of current state-of-the-art methods.