Segmentation and Detection of Pneumothorax using Deep Learning

Pneumothorax is a severe medical condition that can occur when excess air collection in the pleural space between the lung and chest wall. This abnormal build-up of air can cause significant health complications. Early detection of pneumothorax is crucial as it can cause significant harm to the body if not diagnosed promptly. However, even experienced radiologists may miss identifying pneumothorax in chest radiographs. This study proposes a framework that can aid radiologists in quickly and accurately identifying pneumothorax. The framework consists of three steps. First, the chest radiographs are enhanced using histogram equalization to improve the contrast. Second, the lung regions are segmented using a U-Net model to increase the accuracy of classification. Finally, transfer learning approaches such as DenseNet201, DenseNet121, and InceptionV3 are used to classify if the individual is affected or not. Results show that without segmentation, DenseNet121 outperforms the other two approaches. However, with segmentation, DenseNet201 achieves the highest accuracy of 98.5%, outperforming the other two approaches