Why do we use categorical cross entropy for classification of handwritten images into the numbers, instead of binary cross entro

Question

I used both of the loss function, for categorical cross-entropy I got an accuracy of 98.84 % for training data after 5 iterations and with a batch size of 128 on each. For test data, I got 97.92 % accuracy. The code took 24s to execute. But using binary cross-entropy, the accuracy with training data was 99.7 % and that with test data was 99.47% ( smaller difference than the previous one.) The code took 28s to execute.