Quantitative measurement of PNEUMOTHORAX using AI management model and clinical application

Introduction: Delayed or misdiagnosis of pneumothorax may evoke fatal results and artificial intelligence (AI) technique can be a solution. The aim of the study is to devise an AI management model by analyzing images of the patients with pneumothorax, and apply it to the algorithm of experienced thoracic surgeons.

Methods: AI model by deep-learning method was devised using chest radiographs of pneumothorax. U-net was applied for semantic segmentation classifying pneumothorax area and non-pneumothorax area. The segmentation labels for deep learning was produced by an experienced radiologist. The amount of pneumothorax was measured using chest computed tomography (volume ratio; gold standard), chest radiographs (area ratio; true label), and calculated by AI model (area ratio; predicted label). Each value was compared and analyzed with clinical outcomes

Results: Study included 96 patients and 67 were set to be a training and others were test set. AI model for pneumothorax showed that accuracy of 97.8%, sensitivity of 69.2%, and negative predictive value of 99.1%. In a test set, average amount of pneumothorax was 15% in gold standard, 16% in predicted, and 13% in true label. Predicted label was not significantly different from gold standard (p=0.11) but inferior than the true label (difference of MAE; 3.03%p). Amount of pneumothorax for thoracostomy patients were 21.6% in predicted and 18.5% in true.

Conclusions: AI management model of pneumothorax showed high accuracy and negative predictive value. It reflects true amount of pneumothorax and well correlates with real-world practice performed by expert thoracic surgeons. Further research is required to improve sensitivity and practicality