Automated Detection of Pneumothorax using Frontal Chest X-Rays

Abstract

Pneumothorax is the medical term for a collapsed lung. Pneumothorax occurs when air enters the space around the lungs medically termed as the pleural space. Air can find its way into the pleural space when there’s an open injury in the chest wall or a rupture in the lung tissue, disrupting the negative pressure that keeps the lungs inflated. The incidence of pneumothorax was 10% in patients with acute respiratory distress syndrome (ARDS), 24% in patients receiving mechanical ventilation, and 56% in patients requiring invasive mechanical ventilation, with 80% patients died. All 5 patients were male and aged ranging from 54 to 79 years old [1].  The numerous imaging modalities such as standard erect PA chest X-ray, lateral x-rays, expiratory films, supine and lateral decubitus X-rays, thoracic ultrasound scanning, and digital imaging [2-3]. The sensitivity of thoracic ultrasound is found to be 81.8 percent and the specificity is found to be 100 percent. The sensitivity of chest X-ray is found to be 31.8 percent and the specificity is found to be 100 percent [4]. The reports of a study conducted in 2020, had a huge number of patients suspected and admitted for COVID-19 pneumonia. On being examined 0.66% of them developed spontaneous pneumothorax. The study concludes that spontaneous pneumothorax is a rare complication of COVID-19 viral pneumonia and may occur in the absence of mechanical ventilation [5]. Practitioners should be observant towards the development of such complications.

In this paper, automated methods of detecting pneumothorax are explored. Where image segmentation techniques have been employed for detection purposes. The preprocessing methods are handled by image processing techniques using a Support Vector Machine. SVM is a supervised machine learning algorithm that can be used for classification or regression problems. A database of around 10.5k images has been utilized. Initially, the images were preprocessed to remove noise artifacts. The image was then segmented to filter out the region of interest. The images are further passed through the Sobel filter to view the tilts and shifts of the patient’s chest. Furthermore, morphological operations were performed on them to add pixels to image boundaries. This greatly helps in estimating the size of the pneumothorax. Now, in the first phase, the textural features of the X-ray images were extracted. In the second and the last phase, the images were classified based on the severity of pneumothorax.  Resulting in the greater view of the lung in RGB for simplified classification of the normal condition from the pneumothorax affected lung.