Efficacy of Transfer Learning Models in Pneumonia Detection with Scarce Local Healthcare Data

Examining chest X-rays is a difficult and time consuming process even for a professional radiologist. A service that identifies lung issues on X-rays with high accuracy could considerably speed up real world diagnosis processes. In this study, a classification model based on transfer learning was developed using a publicly accessible pneumonia dataset to train the DenseNet-121 model for pneumonia classification. To validate the model’s efficiency on local data, chest X-ray images were collected from different local healthcare centers of Bangladesh. The DenseNet-121 model achieves an accuracy of 95% on the RSNA test dataset and 88% on the local test dataset. This study shows how a model trained on a large publicly available dataset can be used for pneumonia detection on locally collected data despite having differences in region, image format and resolution. This can be very useful when local data is scarce, difficult and time-consuming to collect.