Deep Learning Production In Azure: An Application For Diagnosis In Chest Xrays

Abstract

Lung diseases are the world’s most common diseases, especially in Vietnam. Several thoracic lung diseases can even lead particularly dangerous status for the patients. X-ray is an imaging modality that can be extremely helpful in detecting the abnormalities in the chest area. In addition, artificial intelligence has the ability to empower the detection in X-ray images and to reduce misdiagnosis, knowledge gap between doctors, and pressure on doctors. Therefore, this study aims to apply deep learning technique to detect the abnormalities in chest X-ray images and the data processing methods that are based on data science and statistical approach to improve the performance of deep learning model. The data was explored and processed to obtain the quality data with optimal characteristics. Then, we applied data augmentation, and optimization to the RetinaNet model with ResNet101 in Feature Pyramid Network backbone to achieve the best performance. Our model achieves mean average precision at threshold 0.5 (mAP@0.5) is 0.55 in the validation set with 5 diseases: Aortic enlargement, cardiomegaly, interstitial lung disease, infiltration and nodule/mass. In the other work, we build end-to-end infrastructure for deploying model lifecycle on Azure. We selected Azure service and design solution architecture that system have good performance, high security, high scalability, high availability and it can be applied in the real world.