AI-based 3D segmentation of pulmonary vasculature of CTEPH patients

Chronic thromboembolic pulmonary hypertension (CTEPH) presents challenges for pulmonary artery segmentation due to vascular remodeling, stenosis, and obstructions. This study evaluates a 7-layer dilated convolutional neural network (CNN) with Tversky loss, applied to computed tomography angiography (CTA) images that were preprocessed with image enhancement techniques. The model achieved a Dice score of 0.792 on non-CTEPH data but scored 0.693 on CTEPH data, reflecting the challenges of manual segmentation, where smaller branches are often missed. While the results align with other research, advanced 3D CNN models have shown higher accuracy. Future work should refine ground truth data and explore 3D models to better capture CTEPH-specific complexities.