Regresar

Detection of Pathologies in X-Ray Chest Images using a Deep Convolutional Neural Network with Appropriate Data Augmentation Techniques

Abstract:

Current advances in trained Deep Learning models have allowed architectures, such as Convolutional Neural Networks to outperformed radiologists in developing complex tasks, including detecting pathologies in chest X-ray images. To train these detection models, labeled data sets for identifying pathologies and its bounding box are necessary. However, most available datasets in the scientific community contain the classification labels only, and those with the bounding boxes need to be cleaned and analyzed. In this work, we propose a cleaning methodology and data augmentation techniques that could help perform efficient training procedures, in particular for pathology detection models in chest X-ray images. We found that using these approaches, the models' training Mean Average Precision (mAP) reached 0.93.