Segmentation of Acute Pulmonary Embolism in Computed Tomography Pulmonary Angiography Using the Deep Learning Method

Abstract Background Pulmonary embolism is a type of thromboembolism seen in the main pulmonary artery and its branches. This study aimed to diagnose acute pulmonary embolism using the deep learning method in computed tomographic pulmonary angiography (CTPA) and perform the segmentation of pulmonary embolism data. Methods The CTPA images of patients diagnosed with pulmonary embolism who underwent scheduled imaging were retrospectively evaluated. After data collection, the areas that were diagnosed as embolism in the axial section images were segmented. The dataset was divided into three parts as training, validation, and testing. The results were calculated by selecting 50% as the cut-off value for the intersection over union. Results Images were obtained from 1,550 patients. The mean age of the patients was 64.23 ± 15.45 years. A total of 2,339 axial computed tomography images obtained from the 1,550 patients were used. There were a total of 5,992 labels, with 1,879 images. PyTorch U-Net was used to train 400 epochs, and the best model, epoch 178, was recorded. In the testing group, the number of true positives was determined as 471 and false positives as 35, while 27 were not detected. The sensitivity of CTPA segmentation was 0.95, the precision value was 0.93, the F1 score value was 0.94, and the learning rate was 0.0001. The area under the curve value obtained in the receiver operating characteristic analysis was calculated as 0.88. Conclusions In this study, the segmentation of acute pulmonary embolism in CTPA performed using the deep learning method provided successful results.