Automated pneumothorax segmentation and quantification algorithm based on deep learning

Abstract

A collapsed lung, also known as a pneumothorax, is a medical condition characterized by the presence of air in the chest cavity between the lung and chest wall. A chest radiograph is commonly used to diagnose pneumothorax; however, manual segmentation of the pneumothorax region can be difficult to achieve due to its complicated appearance and the variable quality of the image. To address this, we introduce a two-phase deep learning framework designed to enhance the accuracy of lung and pneumothorax segmentation from chest radiographs. Initially, a U-Net model with a ResNet34 backbone, trained on the Shenzhen and Montgomery datasets, is utilized to achieve precise lung region segmentation. Subsequently, for pneumothorax segmentation, we propose the PTXSeg-Net—a convolutional neural network model trained on the SIIM-ACR pneumothorax dataset. The PTXSeg-Net is an enhancement of the U-Net architecture, modified to incorporate attention gates and residual blocks to refine learning capabilities, further strengthened by deep supervision, allowing for more nuanced gradient utilization across all network layers. We employ transfer learning by pre-training an autoencoder to extract robust chest X-ray representations. Data refinement techniques are applied to the SIIM-ACR dataset to further improve training outcomes. Our results indicate that PTXSeg-Net outperforms other models in pneumothorax segmentation, achieving the highest Dice score of 0.9124 and Jaccard index of 0.8894 on the refined dataset with autoencoder pre-training. Moreover, leveraging the predicted lung and pneumothorax segmentation masks from the two-phase framework, we propose a quantification algorithm for estimating the pneumothorax size ratio. Its validity has been confirmed through expert assessments by a radiologist and a surgeon on a test set comprising 495 images. The high acceptance rates, averaging 96.97 %, demonstrate substantial agreement between the proposed method and expert clinical assessments. The implications of these results are significant for clinical practice, offering a deep learning technology for more accurate and efficient pneumothorax identification and quantification. This improvement facilitates the timely determination of required management and treatment strategies, potentially leading to enhancements in patient outcomes.