Deep Learning based Detection and Segmentation of COVID-19 & Pneumonia on Chest X-ray Image

Abstract

The outbreaks of COVID-19 virus have crossed the limit to our expectation and it breaks all previous records of virus outbreaks. The effect of corona virus causes a serious illness may result in death as a consequence of substantial alveolar damage and progressive respiratory failure. Automatic detection and classification of this virus from chest X-ray image using computer vision technology can be very useful complement with respect to the less sensitive traditional process of detecting COVID-19 i.e. Reverse Transcription Polymerase Chain Reaction (RT-PCR). This automated process offers a great potential to enhance the conventional healthcare tactic for tackling COVID-19 and can mitigate the shortage of trained physicians in remote communities. Again, the segmentation of the infected regions from chest X-ray image can help the medical specialists to view insights of the affected region. So, in this paper we have used deep learning based ensemble model for the classification of COVID-19, pneumonia and normal X-ray image and for segmentation we have used DenseNet based U-Net architecture to segment the affected region. For making the ground truth mask image which is needed for segmenting purpose, we have used Amazon SageMaker Ground Truth Tool to manually crop the activation region (discriminative image regions by which CNN identify a specific class using Grad-CAM algorithm) of the X-ray image. We have found the classification accuracy 99.2% on the available X-ray dataset and 92% average accuracy from the segmentation process.