Pneumonia Classification Using Few-Shot Learning with Visual Explanations

Deep learning models have demonstrated state of the art performance in varied domains, however there is still room for improvement when it comes to learning new concepts from little data. Learning relevant features from a few training samples remains a challenge in machine learning applications. In this study, we propose an automated approach for the classification of Viral, Bacterial, and Fungal pneumonia using chest X-rays on a publicly available dataset. We employ distance learning based Siamese Networks with visual explanations for pneumonia detection. Our results demonstrate remarkable improvement in performance over conventional deep convolutional models with just a few training samples. We exhibit the powerful generalization capability of our model which once trained, effectively predicts new unseen data in the test set. Furthermore, we also illustrate the effectiveness of our model by classifying diseases from the same genus like COVID-19 and SARS.