Adaptive Image Preprocessing and Augmentation for Tuberculosis Screening on Out-of-Domain Chest X-Ray Dataset

Research on deep learning models for chest radiology applications have been getting much attention by the community. These works mostly focus on developing models using in-domain data. However, real-world scenarios might not always match the training set, which makes some models perform worse at the deployment stage. This work focuses on studying the effect of dataset mismatch on chest radiography and methods that can mitigate this problem. We developed the lung balance contrast enhancement technique (lung BCET) which automatically identifies the lung region and normalizes the image accordingly to improve the robustness on out-of-domain data. We also explored augmentation methods that are suitable for chest radiography. We compiled data from multiple tuberculosis (TB) datasets to evaluate and compare the performance of the preprocessing and augmentation methods using the area under the receiver operating characteristic curve (AUC) and heatmap quality. On out-of-domain testing conditions, the lung BCET preprocessing method achieves the highest AUC scores of 0.7978 and 0.6240 for the Maesot and Bureau of TB (BT) datasets, respectively. We also found that lung BCET can also be used to perform data augmentation in conjunction with standard augmentation techniques to improve the performance in both in- and out-of-domain conditions.

Automated summary

Research focuses on the effect of dataset mismatch on chest radiography and proposes lung balance contrast enhancement technique (lung BCET) and augmentation methods suitable for chest radiography. Results show that lung BCET achieves high AUC scores in out-of-domain testing conditions and can be used for data augmentation in both in- and out-of-domain conditions.