Skip to main content

Table 2 Accuracies for the segmentation of a pneumothorax at the pixel level, detection of a pneumothorax at the volume level, and between radiologists, is displayed in terms of sensitivity, specificity calculated at the Youden index of the receiver operating characteristic curve and corresponding area under curve as well as precision, recall, and Dice similarity coefficient calculated at the optimal cut-off point of the precision-recall curve, and corresponding average precision

From: Deep learning detection and quantification of pneumothorax in heterogeneous routine chest computed tomography

Accuracy results

 

Pixel-level segmentation

Inter-rater segmentation

Volume-level detection

Sensitivity

0.958

0.904

0.916

Specificity

0.994

0.999

0.930

Area under the curve

0.979

0.976

Precision

0.961

0.944

0.886

Recall

0.919

0.904

0.880

Dice similarity coefficient

0.939

0.924

0.883

Average precision

0.966

0.954

Number

9

2

567