Hi. Actually, it's pretty easy for a neural network to differentiate between normal and pneumonia x-rays. You will probably get an accuracy of above 95%.

To make the problem a bit difficult, you can separate out the viral and bacterial classes (from pneumonia folder).

If you go into the Pneumonia folder of the train or test sets, you will see that the images are labeled as either bacteria or virus (image file names). That's how i separated them.

Hope this helps.