Choosing the number of neurons in dense layers is something which comes under hyperparameter tuning and the selection of these hyperparameters is heavily dependent on the problem at hand. You’ve to try the wide range of hyperparameters to see which works best for your problem. For this task, 500 in the first dense layer was working fine, that’s why 500 is selected. But the selection of 10 neurons in the final layer is because of 10 class classification problem.

Hope it helps.