Formulating an Integrated Chest Radiograph Dataset to Alleviate Data Scarcity in Respiratory Disease Detection

Respiratory diseases referred to as lung diseases like Pneumonia, Lung-Opacity, Tuberculosis, Lung cancer, Asthma etc. impact the lungs and the different parts of the respiratory system. The primary cause of mortality and disability on a global scale is respiratory diseases. Early diagnosis of the disease aids the medical practitioners and doctors to administer treatment to the patients. We are required to develop an automated computer aided diagnosis (CAD) system for a precise assessment. A major challenge in developing an accurate computer aided diagnosis (CAD) system is the scarcity of large-scale chest radiograph data. Data scarcity signifies constrained access to wide-ranging, extensive and high caliber dataset essential for training convolutional neural network models for a precise disease detection during the preliminary stages. To mitigate the issue of data scarcity in this research, we have acquired and refined three benchmark datasets. Dataset-1 NIH Chest X-rays dataset, Dataset-2 Chest X-Ray Dataset for Respiratory Disease Classification, and Dataset-3 Chest X-Ray Images (Pneumonia) dataset used to formulate an integrated dataset for respiratory disease detection and multiclass classification. In this paper, we have created an imbalanced dataset by implementing data acquisition, data extraction operations on the above benchmark datasets. Enforced data augmentation, data splitting and data pre-processing operations on the formulated imbalanced dataset. The outcome of the implementation is to have vast dataset of chest radiographs, which is categorized into three class labels, Normal (11,775), Other (15,095), and Pneumonia (13,101) classes with approximate size of dataset is 3 GB. The formulated dataset, created by integrating three benchmark datasets, is utilized in the further research work to implement an ensemble model for respiratory disease detection and classification.