Best architecture/optimizer for 3-class prediction?


I am trying to get good scores on a 3-class pneumonia(x-ray images) classification problem which is:
class0 : normal
class1 : bacterial pneumonia
class3 : viral pneumonia

I have tried multiple models(resnet, vgg, etc), pretrained and not, but I cannot surpass 81% testing accuracy.
Is there any model that can perform better on such problem? Or a better optimizer/transformation?

Thank you in advace!