Identification of Pneumothorax from Chest X-ray Images Using Artificial Intelligence Techniques.

Abstract

Technology today has revolutionized the world by replacing the manual system with automatic ones by deploying artificial intelligence, which enables the system to mimic human brain by making wise decisions on the basis of the past experiences. In this research, such computer aided design is proposed which is able to distinguish between pneumothorax and normal X-ray and also solves the class imbalance issue which is troublesome in most of the machine learning classification problems. Such system will help in minimizing the risk of pneumothorax which is a life threatening disease. The proposed CAD system consists of two modules, i.e. classification of chest radiographs as normal or pneumothorax and segmentation model for identifying the location of pathology. For pneumothorax classification, firstly existing approaches for class imbalance are experimented and after finding out that data-level-ensemble outperforms others, an ensemble model is proposed which is actually a model-level-ensemble of multiple data-levelensembles. The different models used in this framework are three different CNN architectures including VGG-16, DenseNet-121 and VGG-19. These architectures are used as fixed feature extractor and support vector machine is used as classifier. The proposed framework is experimented on two datasets: SIM ACR Pneumothorax dataset and Random Sample of NIHChest X-ray dataset (RS-NIH). The model has achieved testing score of 86.0% Area under the Receiver Operating Characteristic curve (AUC) with 85.17% recall on SIIM dataset, while on RSNIH 95.0% AUC with 90.9% Recall is achieved with random split of data and 77.06% AUC with 85.45% Recall is achieved with patient-wise split of the dataset. Our model has performed very well on both the datasets as the AUC achieved on RS-NIH is the best achieved so far, while for SIIM dataset, a direct comparison cannot be made as we are the first to use this dataset for classification. For identification of area of pathology in the CXRs, a two stage segmentation framework is proposed in which the main building block is U-Net architecture with EfficientNetB4 encoder. The images and corresponding masks are resized to 256 x 256 and 384 x 384 for training the first and second stage respectively, using the SIIM dataset. 84.56% dice score is obtained for the segmentation model. Our results prove that the proposed techniques can be generalized to any other medical imagining domain classification and segmentation problem.