DeYOLO: A CNN Based Novel Approach for Classification and Localization of Pneumonia in Chest Radiographs

This paper describes an approach for pneumonia classification and localization of the infected part in the chest radiograph (CXR) images. Localization of pneumonia-infected CXR needs to be in real-time and accurate. Object detection systems like YOLO that are real-time efficient are inaccurate, and there is a trade-off. The proposed DeYOLO model addresses this trade-off by being accurate and real-time with a two-stage detection pipeline. DeYOLO classifies and localizes pneumonia in CXR in real-time with accuracy. DeYOLO is a deep learning model developed using modified neural networks for classification and object detection. Our proposed method, DeYOLO, demonstrates promising experimental results compared with other well-known methods. DeYOLO achieves an ROC score of 0.968 and mAP of 0.209 in the RSNA Pneumonia Detection dataset. DeYOLO can be installed as a web service serving endless users in real-time.