Object Detection for Chest X-ray Image Diagnosis Using Deep Learning with Pseudo Labeling

Object detection has been one of the prominent fields in deep learning. The critical success behind many object detection frameworks is the amount of training data. In some areas, however, the cost of acquisition of labeled data are expensive. In this paper, we propose a method to improve the efficiency of object detection on X-ray images by utilizing the pseudo labeling approach from semi-supervised learning. Our goal is to let the model perform the expert task by labeling the unlabeled samples and train the new model with both newly labeled samples and the original labeled samples. We decide to apply the proposed method to the chest X-ray image diagnosis to replicate the real-world situation in which unlabeled data are easily obtainable compared to labeled ones. From the experimental results, our proposed method improves the mean average precision of the object detection model by 3.56 when the model has 30% of the accessible labels and 5.67 when the 50% of labeled data are accessible.