Differentiation of bacterial and viral pneumonia in children under five using deep learning

With 92,000 deaths and 18 percent of the total child mortality every year, pneumonia is the leading cause of child mortality in children under 5 in Pakistan. Pakistan is one of the top 5 countries for childhood pneumonia deaths around the world. Bacteria and viruses are most common infectious agents of pneumonia. The diagnostic test for pneumonia detection is chest x-ray. Basic diagnostic tests facilities are available even at rural health centers. In proposed study, a pre-trained convolutional neural network; VGG19 model is fine-tuned on dataset of 5863 chest x-ray images of healthy, viral, and bacterial pneumonia. The VGG19, model 1 is trained on viral and bacterial pneumonia images, and model 2 is trained on multi-class data. The model 1 with viral and bacterial pneumonia images showed training accuracy of 0.83 and validation accuracy of 0.84. The model 2 with normal, viral, and bacterial pneumonia, showed training accuracy of 0.84 and validation accuracy of 0.85. The results show that the VGG19 model has powerful prediction capacity to identify correct features of types of pneumonia with reasonable accuracy even with smaller and unbalanced dataset. The results predict that already developed and trained algorithms can be used as ready to use clinical diagnostic tool, if fine-tuned with larger balanced dataset with few targeted changes. These tools can be used as second reader tool by the physicians, can process thousands of images in limited time with high accuracy, relieving the burden of patients on limited capacity of the healthcare facilities.