PneuNetV1: A Deep Neural Network for Classification of Pneumothorax Using CXR Images

Abstract

Pneumothorax is a critical medical condition among human-beings. A severe pneumothorax causes collapsed lungs. It is a life-threatening disease. Therefore, pneumothorax detection is an important step for the prevention and curing of a patient. Pneumothorax can be classified into three major categories: primary, secondary, and injury. Magnetic resonance imaging (MRI)-based digital imaging and communications in medicine (DICOM) files of Chest X-ray (CXR) images provide insight and help the doctor to make an appropriate decision. An early decision can prevent the mortality rate among patients. Since the outbreak of the COVID-19 pandemic, the medical systems and staff have gone under massive pressure. Classification from a CXR image by an expert requires huge manpower and a longer time to determine. Deep learning-based automatic classification of Pneumothorax (CXR) images can assist the medical community in a fast diagnosis and reduce the burden of work overload. Doctors can focus on better treatment and cure of Pneumothorax. In this paper, we have proposed seven scratch Convolutional Neural Networks (CNN) architectures and compared them with another seven transfer learning models. The best-performing CNN model (PneuNetV1) is determined based on various standard performance metrics. It has gained the highest test accuracy, efficacy ratio, and F1-score of 0.9123, 5.2370, and 0.9220, respectively with a minimum training time. The obtained results are achieved through rigorous experimentation and yet provide satisfactory performance.

Keywords

• Convolutional neural network (CNN)
• CXR images
• pneumothorax
• transfer learning
• VGG-19