Segmentation of Pneumothorax on Chest CTs Using Deep Learning Based on Unet-Resnet-50 Convolutional Neural Network Structure

Abstract

Objective: Pneumothorax refers to an abnormal accumulation of air in the pleural cavity. This condition is significant in terms of health and can provide a life-threatening risk, particularly when it is extensive or occurs alongside other medical conditions. Nevertheless, the scarcity of work on chest CT segmentation arises from the challenge of acquiring pixel-level annotations for chest X-rays. This paper presents and assesses a deep learning approach utilizing the Unet-Resnet-50 convolutional neural network architecture for accurately segmenting pneumothoraces on chest computed tomography (CT) images.

Methods: We employed a private dataset including 2627 manually annotated slices obtained from 16 patients. We assessed the model's performance by measuring the dice similarity coefficient (DSC or F1 score), accuracy, area under the curve (AUC), precision, and recall on both the validation and test sets.

Results: The binary accuracy of the test set was 0.9990; the precision was 0.9681; and the DSC was 0.9644. Although it contains less data (16 patients), we found that our deep learning-based artificial intelligence model has effective and compatible results with the literature.

Conclusion: Deep learning models that will be used to detect common pathologies in thoracic surgery practice, such as pneumothorax, to determine their localization and size, will provide faster diagnosis and treatment to patients, and especially improve radiology workflow.