The Effect of PreTraining Thoracic Disease Detection Systems on Large-Scale Chest X-Ray Domain Datasets

Abstract

The COVID-19 pandemic has impacted many countries around the world resulting in the need to develop quick and effective screening methods to ease the burden and overcome the limitations of varying healthcare capacities. Given the nature of the disease, the use of Chest X-ray (CXR) medical imaging has proven to be very useful which has prompted the exploration of computer-aided diagnosis tools to augment and assist radiologists. However, recent reports have deemed many of the proposed methods to be impractical for use in real-life applications due to models with poor generalization capabilities, an issue closely related to the quality of current datasets in the CXR domain. Typically, deep convolutional neural network (CNN) based classification systems utilize transfer learning techniques when data is limited. We suggest first training models on publicly available large-scale and CXR specific datasets, such as CheXpert, and using these pretrained weights when initializing the final model. Compared with a CNN pretrained on the more general ImageNet dataset, pretraining on large-scale domain specific data increased the model's ability to generalize to unseen data.