Application and Challenges of Deep Learning in Pulmonary Vessels Segmentation of CTPA Images.

Abstract

Accurate segmentation of pulmonary vessels in medical imaging is critical for the diagnosis of pulmonary vascular diseases (PVDs), particularly in conditions such as chronic thromboembolic pulmonary hypertension (CTEPH), which require detailed vascular mapping. This comprehensive review explores recent advancements in deep learning (DL)-based segmentation techniques for computed tomography pulmonary angiography (CTPA) images, focusing on three primary objectives: (1) to systematically classify network architectures by data dimensionality (2D/3D/2.5D) and assess their clinical adaptability across varying imaging conditions; (2) to perform quantitative performance comparisons using the standardized Dice Similarity Coefficient (DSC); and (3) to address critical challenges in clinical implementation, including annotation scarcity, computational efficiency versus resolution trade-offs, and model generalization limitations, while proposing innovative mitigation strategies. To this end, we adopted the PRISMA methodology, conducting rigorous searches across Google Scholar and IEEE Xplore databases up to 2024, with manual curation identifying 23 high-quality studies for inclusion. In addition to synthesizing current technological limitations, this review highlights emerging directions such as self-supervised learning and domain adaptation, offering clinicians and AI researchers a structured reference that bridges technical innovation with practical clinical needs.