Deep learning pulmonary vessel segmentation for contrast-enhanced and non-contrast CT images.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China.
- Research and Development Center, Canon Medical Systems (China), Beijing, China.
Abstract
Accurate and comprehensive segmentation of pulmonary vasculature is essential for diagnosing various pulmonary diseases and lung surgical planning. To develop a deep learning model for pulmonary vessels segmentation based on both computed tomography pulmonary angiography (CTPA) and non-contrast computed tomography (NCCT) data, and to evaluate its performance on a clinical dataset. In this retrospective study, we collected 427 cases of chest CT data from different countries and manufacturers, including 213 CTPA cases and 214 NCCT cases. And we designed a CTPA and NCCT ground truth (GT) generation method producing high-accuracy vessel annotation dataset. Then we developed a 3D U-Net based deep learning method with the Vessel Lumen Structure Optimization Module (VLSOM). We numerically evaluated model using metrics including Centerline Recall (Cl-Recall), Centerline DICE (Cl-DICE), Average Symmetric Surface Distance (ASSD) and 95% Hausdorff Distance (HD95) focusing on the integrity of vascular structures. Finally, the model performance was visually assessed in an independent external dataset. In the numerical evaluation on the internal test dataset with 57 cases, our model achieved Cl-DICE, Cl-Recall and HD95 values of 0.909 (95% CI [0.904, 0.914]), 0.883 (95% CI [0.873, 0.891]) and 1.861 mm (95% CI [1.674 mm, 2.073 mm]) which is superior to the state of the art (SOTA) baseline model nnU-Net v2 of 0.860 (95% CI [0.852, 0.867]), 0.772 (95% CI [0.760, 0.783])and 2.458 mm (95% CI [2.219 mm, 2.708 mm]) (all p < 0.001). In the clinical visual assessments on the 90 sets of external datasets, the average score for accuracy and robustness, branch abundance, computer-aided detection and vascular continuity 4.52 (95% CI [4.42, 4.62]), 4.82 (95% CI [4.76, 4.90]), 4.58 (95% CI [4.46, 4.69]), and 4.85 (95% CI [4.77, 4.92]) respectively with a full score of 5. The proposed pulmonary vessel segmentation algorithm performed well in preserving the connectivity and integrity of the vascular structure on both CTPA and NCCT data with various diseases. A vessel centerline topology-enhanced deep segmentation model enables accurate and automatic pulmonary vessel segmentation in both computed tomography pulmonary angiography (CTPA) and non-contrast computed tomography (NCCT) images.