Deep learning-enabled system for rapid pneumothorax screening on chest CT.

PURPOSE: Prompt diagnosis and quantitation of pneumothorax impact decisions pertaining to patient management. The purpose of our study was to develop and evaluate the accuracy of a deep learning (DL)-based image classification program for detection of pneumothorax on chest CT.

METHOD: In an IRB approved study, an eight-layer convolutional neural network (CNN) using constant-size (36*36 pixels) 2D image patches was trained on a set of 80 chest CTs, with (n = 50) and without (n = 30) pneumothorax. Image patches were classified based on their probability of representing pneumothorax with subsequent generation of 3D heat-maps. The heat maps were further defined to include 1) pneumothorax area size, 2) relative location of the region to the lung boundary, and 3) a shape descriptor based on regional anisotropy. A support vector machine (SVM) was trained for classification.

RESULT: We assessed performance of our program in a separate test dataset of 200 chest CT examinations, with (160/200, 75%) and without (40/200, 25%) pneumothorax. Data were analyzed to determine the accuracy, sensitivity, specificity. The subject-wise sensitivity was 100% (all 160/160 pneumothoraces detected) and specificity was 82.5% (33 true negative/40). False positive classifications were primarily related to emphysema and/or artifacts in the test images.

CONCLUSION

This deep learning-based program demonstrated high accuracy for automatic detection of pneumothorax on chest CTs. By implementing it on a high-performance computing platform and integrating the domain knowledge of radiologists into the analytics framework, our method can be used to rapidly pre-screen large numbers of cases for presence of pneumothorax, a critical finding.