Test set has low accuracy

I'm training a neural network with 4-fold cross-validation and I reach over 90% accuracy on validation set, but when I test it, accuracy is 50% and I don't know why. 80% of the dataset is used for training and 20% for testing.