Validation accuracy and prediction accuracy is different

Question

I am doing the image classification task for grading the gradability of the fundus images with 3 classes: Good, Usable and Bad. The problem is during the training, the valid