This paper presents, methodical comparisons between four CNN architectures and different learning approaches, for detecting pneumonia in X-Ray images. We evaluate 12 different models obtained by applying three different learning approaches on four different CNN architectures. The results show that transfer learning using fine-tuning performs quite well on all cnn architectures, showing little or no overfitting in most cases. For the overall top model, we find that ResNeXt-50 with fine tuning performs the best. Achieving a high sensitivity (recall) of 98.7%, 75.6% specificity and AUROC of 0.87.