Automatic pulmonary vessel network labeling on thoracic CT scans based on partially labeled data

Révy, Gábor

Hadházi, Dániel

Hullám, Gábor

2025-05-22T11:43:57Z

2025-05-22T11:43:57Z

2025-05-23

Accurate pulmonary vessel segmentation, and the separation into arterial and venous networks helps doctors by guiding them during surgical planning, particularly in thoracic procedures. Despite significant advancements, the automation of this task is still currently in progress. Many studies show promising results, yet they often rely on high-quality CT scans, which are not always available in routine clinical practice. Specifically, these algorithms typically depend on contrast-enhanced CT scans with thin slice spacing, precise timing of contrast injection, and the application of filtering algorithms during reconstruction to minimize streaking artifacts. In our previous work we presented a set of algorithms to improve an existing system for artery-vein separation, with the aim of making it more robust on CT scans of typical quality. Despite evaluating a wide variety of algorithms, it became clear that an explicit, model-based approach was necessary to achieve the desired accuracy and reliability in artery-vein separation. However, developing such a model requires a substantial amount of properly labeled data, which is difficult to obtain due to the time-consuming nature of manual labeling. We investigated the dataset of the PARSE2022 (Pulmonary Artery Segmentation) challenge with incomplete annotations: only arteries are segmented on the CT scans. This poses a considerable obstacle for training a robust model. To address this, we developed an algorithmic solution based on bilateral filtering to automatically complete the segmentation with the venous labels, ensuring that the dataset was sufficiently annotated for model training.

angol

könyvfejezet

application/pdf

http://hdl.handle.net/10890/60574