Transfer Learning based Noise Classification in Chest X-Ray Images

The chest radiograph (Chest X-ray) is the most commonly requested radiological examination owing to its effectiveness in the characterization and detection of cardiothoracic and pulmonary abnormalities. Deep Learning is emerging as the leading tool for medical image analysis. In particular, deep Convolutional Neural Networks (CNNs) have proven to be powerful tools for a wide range of computer vision tasks. The presence of noise in medical images reduces the efficacy of disease diagnosis and detection using manual inspection. This paper discusses different types of noise typically detected in Chest radiographs. The dataset is created using synthesized noise added mathematically onto the images. Transfer Learning methods are implemented in CNN based architectures to determine the best suitable model and the accuracy for real-time implementation. A comparison is also performed in training the model using data with augmentation and training without data augmentation. The geometrical variations induced in data augmentation and the three levels of noise added for each category suffice the requirement for generalizing the model to obtain validation accuracy of 99.41% for VGGnet and DenseNet. MobileNet V2 having inverted residual structure performs well on mobile devices that achieved an accuracy of 96.17%.