A Self-training Framework for Semi-supervised Pulmonary Vessel Segmentation and Its Application in COPD

Shuiqing Zhao, Meihuan Wang, Jiaxuan Xu, Jie Feng, Wei Qian, Rongchang Chen, Zhenyu Liang, Shouliang Qi, Yanan WuarXiv 2025 · DOI · arXiv

Abstract

Accurate segmentation of the pulmonary vasculature is essential for quantifying disease progression in Chronic Obstructive Pulmonary Disease (COPD). However, manual annotation is labor-intensive, and automated methods struggle with the low contrast of peripheral vessels and disease-induced artifacts. We present a semi-supervised self-training framework that combines a hybrid CNN-Transformer architecture (nnFormer) with a teacher-student learning loop. Our approach achieves a 90.3% precision rate on a dataset of 125 COPD CT scans. Clinical application reveals quantitative evidence of "vessel pruning," where small vessel volume decreases relative to total blood volume as disease severity increases.

Introduction

Chronic Obstructive Pulmonary Disease (COPD) profoundly affects the pulmonary vasculature, leading to structural changes that are critical for understanding disease pathophysiology. Segmenting these vessels—particularly the small, peripheral ones—is a significant challenge for computer vision due to their intricate shapes, low contrast, and the presence of emphysema-related noise.

Deep learning models typically require large, high-quality labeled datasets, which are scarce for 3D pulmonary vessel segmentation. This study addresses the "annotation bottleneck" through a semi-supervised framework that leverages a small labeled dataset alongside a larger pool of unlabeled images.

Methodology

The proposed pipeline utilizes a multi-stage self-training approach to maximize the utility of limited expert annotations.

Interactive Annotation and Backbone Architecture

To establish a high-quality "gold standard," 12 cases were initially segmented using coarse models and subsequently refined by radiologists. The core architecture employed is nnFormer, a hybrid model that integrates convolutional neural networks (CNNs) for local feature extraction with Transformers for modeling long-range dependencies. This dual capability is vital for capturing both the fine texture of individual vessels and the global context of the vascular tree.

Detailed segmentation of the pulmonary vascular tree, highlighting the model's ability to capture fine peripheral branches.

Teacher-Student Self-Training

The framework operates via a teacher-student loop:

  1. Teacher Training: A teacher model is trained on the small, expert-labeled dataset.
  2. Pseudo-labeling: The teacher generates initial masks for the unlabeled data.
  3. Selection Strategy: Only the most reliable pseudo-labels, determined by high precision and Dice scores, are retained.
  4. Student Training: A student model is trained on the combined set of original labels and high-quality pseudo-labels.

To further improve accuracy at vessel boundaries and terminations, a Hard Region Adaptation (HRA) Loss was implemented, which dynamically increases the penalty for errors in these challenging areas.

Datasets

The study utilized an in-house dataset of 125 non-enhanced CT scans from COPD patients, categorized by the GOLD (Global Initiative for Chronic Obstructive Lung Disease) system (GOLD 1-4). Twelve cases were used for training, 10 for testing, and 103 served as unlabeled data for semi-supervised learning. The VESSEL12 dataset was used as an external reference for initial model pre-training.

Results and Clinical Insights

The semi-supervised approach (Semi2) achieved a precision of 90.3%, a 2.3% improvement over the fully supervised baseline. In comparative tests, nnFormer (Dice: 0.855) outperformed other architectures, including UNETR and nnU-Net.

Comparison of segmentation performance in lungs with varying degrees of emphysema, illustrating the model's robustness and the 'vessel pruning' effect.

A key clinical finding was the quantification of the "vessel pruning" phenomenon. While total blood volume (TBV) increased with disease severity—likely due to inflammation—the volume of the smallest vessels (BV5) decreased. Specifically, the BV5/TBV ratio dropped from 0.570 in GOLD 1 patients to 0.521 in GOLD 4 patients, providing a robust quantitative metric for disease progression.

Discussion

While the framework demonstrates high precision and clinical utility, limitations include the use of single-center data. Future work will focus on multi-center validation to ensure robustness across different imaging protocols. Additionally, integrating vascular metrics with other clinical indicators, such as lung function tests (FEV1/FVC), could provide a more comprehensive tool for COPD management. The exploration of generative models, such as Diffusion Models, may further enhance the segmentation of micro-vessels in future iterations.

Summary updated

BibTeX
@misc{zhao2025self,
  author = {Shuiqing Zhao and Meihuan Wang and Jiaxuan Xu and Jie Feng and Wei Qian and Rongchang Chen and Zhenyu Liang and Shouliang Qi and Yanan Wu},
  title = {{A Self-training Framework for Semi-supervised Pulmonary Vessel Segmentation and Its Application in COPD}},
  year = {2025},
  eprint = {2507.19074},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.19074},
}
Figure