The interpretation of chest x-rays is critical for the discovery of thoracic diseases, including pneumonia and lung cancer, which affect millions of people worldwide each year. This time-consuming task usually requires radiologists to read the images, leading to diagnostic errors due to fatigue and lack of diagnostic expertise in areas where there are no radiologists in the world. Recently, deep learning methods have been able to perform well in the field of medical imaging, thanks to the emergence of large network architectures and large labeled datasets. In this work, we describe our approach to pneumonia classification and localization in chest radiographs. This method uses only open-source deep learning object detection and is based on RetinaNet, a fully convolutional network which incorporated global and local features for object detection. Our method achieves the classification and localization of Chest radiograph pneumonia by key modifications to the image preprocessing and training process, and incorporates bounding boxes from multiple models during the test. Improve the effect of algorithm classification and localization. After image enhancement and algorithm improvement, we randomly selected 100 chest radiographs on the second stage chest dataset to test our detection algorithm and achieved good results. Our findings yield an accuracy of 90.25%.
Paper
Full text
Pneumonia detection based on deep neural network Retinanet
Semantic Scholar · Medicine · 2019
Abstract
The interpretation of chest x-rays is critical for the discovery of thoracic diseases, including pneumonia and lung cancer, which affect millions of people worldwide each year. This time-consuming task usually requires radiologists to read the images, leading to diagnostic errors due to fatigue and lack of diagnostic expertise in areas where there are no radiologists in the world. Recently, deep learning methods have been able to perform well in the field of medical imaging, thanks to the emergence of large network architectures and large labeled datasets. In this work, we describe our approach to pneumonia classification and localization in chest radiographs. This method uses only open-source deep learning object detection and is based on RetinaNet, a fully convolutional network which incorporated global and local features for object detection. Our method achieves the classification and localization of Chest radiograph pneumonia by key modifications to the image preprocessing and training process, and incorporates bounding boxes from multiple models during the test. Improve the effect of algorithm classification and localization. After image enhancement and algorithm improvement, we randomly selected 100 chest radiographs on the second stage chest dataset to test our detection algorithm and achieved good results. Our findings yield an accuracy of 90.25%.