Pulmonary vessel segmentation in computed tomography images: a cascaded approach combining U-Net and parameter-adaptive fully connected conditional random fields

Abstract

Background: A precise pulmonary vessel segmentation algorithm serves as a powerful auxiliary tool for physicians, enabling them to diagnose various pulmonary diseases with greater accuracy and efficiency. This technology also customizes rational treatment plans tailored to individual patients, alleviating their burden and effectively reducing unnecessary medical resource waste. This study proposes a cascaded algorithm to improve the accuracy of pulmonary vessel segmentation in computed tomography (CT) images.

Methods: This study presents a cascaded model integrating convolutional networks for biomedical image segmentation (U-Net) and parameter-adaptive fully connected conditional random fields (PA-FCCRFs) to efficiently extract pulmonary vessels in CT images. In the initial phase, U-Net is employed to preliminarily segment pulmonary vessels in the lung region. However, convolutional neural network (CNN) with local receptive fields struggles to effectively model long-distance pixel dependencies, often leading to mis-segmentation of lung tissues. To address this issue, we incorporate fully connected conditional random fields (FCCRFs) into the framework for refined segmentation. With fully connected structure, FCCRFs can model dependencies between each pixel and all the other pixels. Moreover, Bayesian optimization is employed to automatically tune internal parameters for optimal performance.

Results: Our method demonstrates significant improvements in pulmonary vessel segmentation outcomes, with the Precision increasing from 73.14±10.67 to 90.24±4.63, F1 improving from 82.67±6.86 to 91.85±3.41, and Hausdorff distance decreasing from 35.12±6.04 to 30.86±2.71. To validate the cascaded PA-FCCRFs strategy, we preliminarily segment pulmonary vessels using AH-Net and V-Net, followed by optimization using PA-FCCRFs. Experimental results showcase substantial enhancements in the accuracy of CNN-based vascular segmentation after PA-FCCRFs optimization.

Conclusions: These findings validate that the cascaded PA-FCCRFs approach effectively segments pulmonary vessels, supporting the diagnosis of pulmonary diseases and promising applications in clinical settings.

Keywords: Bayesian optimization; Pulmonary vessel segmentation; convolutional networks for biomedical image segmentation (U-Net); fully connected conditional random fields (FCCRFs).