Location Fusion and Data Augmentation for Thoracic Abnormalites Detection in Chest X-Ray Images
The application of deep learning in medical image diagnosis has been widely studied recently. Unlike general objects, thoracic abnormalities in chest X-ray radiographs are much harder to label consistently by domain experts. The
problem’s difficulty and inconsistency in data labeling lead to the downgraded performance of the robust deep learning models. This paper presents two methods to improve the accuracy of thoracic abnormalities detection in chest X-ray images. The first method is to fuse the locations of the same abnormality marked differently by radiologists. The second method is applying mosaic data augmentation in the training process to enrich the training data. Experiments on the VinDr-CXR chest X-ray data show that combining the two methods helps improve the predictive performance by up to 8% for F1-score and 9% for the mean average precision (MAP) score.