Artificial intelligence studies are increasing day by day in all fields of science, especially medicine. As a current issue, in our study, we segmented pulmonary embolism using the deep learning method on axial section images of patients with pulmonary artery embolism who underwent CTPA. In the testing group of our study, the sensitivity, precision, F1 score and AUC values obtained with our artificial intelligence model were measured as 0.95, 0.93, 0.94, and 0.88, respectively, indicating successful results in the diagnosis and segmentation of pulmonary embolism.
Weikert et al. evaluated the performance of an artificial intelligence algorithm called AI-powered algorithm and detected pulmonary embolism on CTPA images. The authors used approximately 28,000 CTPA images for validation, and utilized the ResNet architecture in the convolutional neural network application. The sensitivity value of the AI-powered algorithm in the detection of pulmonary embolism was found to be 92.7% [18]. In contrast, in our study, segmentation was performed instead of detection. Our sensitivity value was 95%, which is slightly better compared to the value reported by Weikert et al. In addition, we did not include the images of patients without pulmonary embolism, unlike the previous study [18].
In another study using the computer-aided detection algorithm, the sensitivity was calculated separately for the detection of emboli in the main pulmonary artery, lobar, segmental and subsegmental arteries, and as expected, the sensitivity in the detection of embolism in the main pulmonary artery (87%) was found to be higher compared to the subsegmental artery (61%) [19]. Different from this study, we did not perform separate calculations for embolisms detected in the main pulmonary, lobar, segmental and subsegmental arteries.
Rucco et al. used the neural hypernetwork for the diagnosis of pulmonary embolism and obtained from the images of 1,427 patients. This method successfully diagnosed pulmonary embolism at a rate of 94% [20], which is quite similar to our study. In another study conducted using the deep learning method, multimodal fusion was used, and both clinical and laboratory data and images of the patients were evaluated [21]. The AUC value of that study was higher than we obtained. This may be due to the previous authors’ inclusion of clinical and laboratory data in their evaluation.
Huang et al. used the PENet system and attempted to diagnose pulmonary embolism on volumetric CT images with 3D CNN in the infrastructure [22]. The authors found the AUC value to be approximately 0.85, indicating that our method was more successful. In addition, different from our study, Huang et al. performed detection rather than segmentation.
In another study that aimed to detect pulmonary embolism with deep learning, clot burden assessment was also performed, and segmentation was applied [23]. Unlike our study, the authors also calculated the volume of embolism and measured cardiovascular parameters on CT for pulmonary embolism.
We consider that our study is clinically important because the method presented shortens the patient service and evaluation time. In addition, the workload of radiology departments can be reduced using the deep learning-based segmentation model we created, and this will contribute to this field.
One of the limitations of our study is that the segmental branches of the pulmonary artery and the main pulmonary arteries were not considered separately. In addition, the success of the model presented in our study can be increased by including the clinical and laboratory findings of the patients and evaluating cardiovascular parameters that contribute to the course of pulmonary embolism using CT. This will help obtain more successful model alternatives.