Evaluation of transfer learning in Pneumonia classification

Abstract

Pneumonia is a major cause of hospitalization and death in both children and adults. An early diagnosis of pneumonia using low-cost chest X-rays can help improve the chance of survival. However, analyzing the X-rays requires expert medical professionals as resources that might not be available in under-developed regions. Computer-aided chest X-ray diagnosis can help empower and augment clinical decision making in such cases. For image classification tasks, Deep Transfer Learning is proven to be fast and effective. In this work, we use the pre-trained DenseNet architecture as the base and train a classifier to recognize pneumonia in chest X-ray images. We finetune the DenseNet blocks at different levels of the architecture in our experiments and study the impact on the classification. The results reveal that finetuning the lower layers is not as impactful as finetuning the top layers.