Real time Chest X-ray Pathology detection and localization framework with Convolutional Neural Networks and Ensembling

The primary methods of detection of abnormalities in the chest and heart region are analysis of Chest X-rays by expert radiologists trained over multiple years. The cycle time and accuracy for diagnosis with radiographs are often not favorable for the patients leading to undesirable outcomes. To mitigate this issue, we have developed a fully online end-to-end radiology application that performs highly reliable classification and localization of 14 different pathologies using an ensemble of different deep convolutional Neural network architectures and presents it as an editable output for any radiologist to investigate, make changes and confirm the outcome. Our core Deep learning algorithm also achieves a meaningful mAP@40 on VinDr-CXR test set. This is very significant achievement and can help to reduce subjectivity in detection by bringing in more consistency.