Segmentation and Measurements of Major Blood Vessels from CT Pulmonary Angiography Using Deep Learning

BackgroundMeasuring major blood vessels from CT pulmonary angiography examinations (CTPAs) to assess cardiovascular diseases has the potential to improve overall patient outcomes. However, this process is time-consuming and prone to errors. Deep learning (DL) approaches offer the potential to enhance accuracy, speed, and consistency. ObjectivesTo develop and train a deep learning-based algorithm capable of automatically and accurately segmenting and measuring major blood vessels in CTPAs. MethodsSeven hundred CTPAs from 652 patients were retrospectively collected at a single center. The dataset was split into two subsets, one for training and cross-validation (n = 490) and one for assessing model performance (n = 210). The segmentation masks for the descending aorta (DAo), ascending aorta (AAo), and pulmonary trunk (PT) were generated by our previously developed segmentation model and were quantitatively validated by two radiologists. These validated masks were subsequently used as ground truth for model training. An U-Net deep learning model was created using the nnU-Net framework and trained on 490 CTPAs with 5-fold cross-validation. Following the training, the model was applied to volumes of interest in the images to generate a pool of candidate regions containing potential vessels. A vessel detection algorithm was developed and used on the candidate pool to identify vessels followed by measurement. The final model was evaluated on 210 and 47 CTPAs from internal and external datasets, respectively. ResultsAssessing model segmentation performance on the internal evaluation set, the median Dice scores were 0.95 for the DAo, 0.96 for the AAo, and 0.95 for the PT. The model measurements showed a strong correlation with those made by the radiologist, with Pearsons r values of 0.91 for image noise, 0.98 for intravenous contrast concentration in the PT, 0.93 for AAo diameter, and 0.55 for PT diameter (P < .001). Additionally, the AAo segmentation (median Dice score 0.94) and the PT diameter measurement (r = 0.77) were evaluated in two external datasets. ConclusionsThe fully automated, deep learning-based algorithm accurately segmented and measured major blood vessels in real-world CTPAs.