Combating Bias in COVID-19 Disease Detection Using Synthetic Annotations on Chest X-Ray Images

Detecting a COVID-19 case by using a deep learning model poses a challenge due to the use of public datasets, where people can contribute and submit images without quality screenings. One of the challenges is that we found many images contain burned-in annotations, such as tubes, letters, numbers, pads, arrows, etc. The annotations become more problematic if multiple datasets are combined due to the limited number of data for COVID-19 cases, and the other datasets do not contain as many burned-in annotations as in datasets containing samples for COVID-19 cases. An example of the issues is that by using a saliency map method, we can find the troubled areas coincide with areas where the annotations are located. In order to combat this annotation bias, we investigate the effect of deliberately adding synthetic annotations to images for all classification classes. Encouraging results are shown in this paper. That is, by using the proposed method, the F1-score can be improved, e.g., an improvement of F1-score of 0.88 can be increased up to 0.94. Therefore, we conclude that adding synthetic annotations in the pre-processing pipeline for datasets having annotation bias could improve a machine learning model.