High Accuracy Pulmonary Vessel Segmentation for Contrast and Non-contrast CT Images and Clinical Evaluation

summary

The gist

Accurate segmentation of pulmonary vessels plays a very critical role in diagnosing and assessing various lung diseases<ref:2503.16988#pg2> This study proposes a 3D image segmentation algorithm for…

In short

This study developed a 3D image segmentation algorithm for automatically identifying pulmonary vessels from both contrast-enhanced and non-contrast CT scans. The model uses a Vessel Lumen Structure Optimization Module (VLSOM) to extract vessel centerlines and apply specialized loss functions. The resulting model achieved high accuracy across various metrics, outperforming existing methods.

Key concepts

3D U-Net Architecture
This is the core structure of the deep learning model used for segmentation. It's an encoder-decoder network with skip connections, which helps the model capture both broad context and fine details in complex medical images like CT scans. This design is highly effective for accurately delineating structures in 3D space.
Vessel Lumen Structure Optimization Module (VLSOM)
This specialized module enhances the standard 3D U-Net by focusing on vessel centerlines (Cl). It extracts these centerlines from ground truth data and uses them to guide the training process, applying weighted supervision based on how close a vessel segment is to its true centerline. This helps stabilize the segmentation of vessels.
Cl-DICE Loss
This is a specific loss function added to the model's training process. It supervises the stability and structure of the segmented vessels by incorporating information about their extracted centerlines (Cl). By minimizing this loss, the model learns to generate vessel segments that not only match the overall shape but also maintain a realistic and continuous vascular structure.

Terminology used across episodes

This episode discusses

The paper

High Accuracy Pulmonary Vessel Segmentation for Contrast and Non-contrast CT Images and Clinical Evaluation · Read on arXiv

Ying Ming, Shaoze Luo, Longfei Zhao, Ruijie Zhao, Bing Li, Qiqi Xu, *Wei Song

Department of Radiology, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100730, China · Research and Development Center, Canon Medical Systems (China), No.3, Xinyuan South Road, Chaoyang District, Beijing 100027

Accurate segmentation of pulmonary vessels plays a very critical role in diagnosing and assessing various lung diseases. Currently, many automated algorithms are primarily targeted at CTPA (Computed Tomography Pulmonary Angiography) types of data. However, the segmentation precision of these methods is insufficient, and support for NCCT (Non-Contrast Computed Tomography) types of data is also a requirement in some clinical scenarios. In this study, we propose a 3D image segmentation algorithm for automated pulmonary vessel segmentation from both contrast-enhanced and non-contrast CT images. In the network, we designed a Vessel Lumen Structure Optimization Module (VLSOM), which extracts the centerline (Cl) of vessels and adjusts the weights based on the positional information and adds a Cl-Dice Loss to supervise the stability of the vessels structure. We used 427 sets of high-precision annotated CT data from multiple vendors and countries to train the model and achieved Cl-DICE, Cl-Recall, and Recall values of 0.892, 0.861, 0.924 for CTPA data and 0.925, 0.903, 0.949 for NCCT data. This shows that our model has achieved good performance in both accuracy and completeness of pulmonary vessel segmentation. We finally conducted a clinical visual assessment on an independent external test dataset. The average score for accuracy and robustness, branch abundance, assistance for diagnosis and vascular continuity are 4.26, 4.17, 4.33, 3.83 respectively while the full score is 5. These results highlight the great potential of this method in clinical application.

DOI: 10.1002/mp.70751

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "High Accuracy Pulmonary Vessel Segmentation for Contrast and Non-contrast CT Images and Clinical Evaluation".

Dev: Here's the gist of the paper. Accurate segmentation of pulmonary vessels plays a very critical role in diagnosing and assessing various lung diseases<ref:2503.16988#pg2> This study proposes a 3D image segmentation algorithm for automated pulmonary vessel segmentation from both contrast-enhanced and non-contrast CT images<ref:2503.16988#pg2> The network designed includes a Vessel Lumen Structure Optimization Module (VLSOM), which extracts the centerline (Cl) of vessels and adjusts the weights based on the positional information and adds a Cl-Dice Loss to supervise the stability of the vessels structure<ref:2503.16988#pg2> The model achieved Cl-DICE, Cl-Recall, and Recall values of 0.892, 0.861, 0.924 for CTPA data and 0.925, 0.903, 0.949 for NCCT data<ref:2503.16988#pg2> Data and Annotation The dataset consisted of two main sources: 112 cases from the publicly available LIDC (Lung Image Database Consortium) dataset and 315 cases from multiple clinical institutions with the approval of the institutional review boards<ref:2503.16988#pg4> The dataset included 213 cases of CTPA and 214 cases of NCCT cases, which facilitated the development of a versatile model capable of handling both contrast-enhanced and non-contrast imaging data<ref:2503.16988#pg5> Ground truth (GT) annotations were performed by two radiologists with more than ten years of working experience to ensure high-quality supervision during model training<ref:2503.16988#pg6> The annotation standard required labeling vessels as small as 1 mm in diameter, ensuring that even fine vascular structures were captured<ref:2503.16988#pg7> A data annotation scheme was designed to simultaneously obtain high-precision GT results for both CTPA and NCCT, enabling the trained segmentation model to support automatic segmentation for both types of data<ref:2503.16988#pg8> Model Architecture and Loss Function The backbone of the model is based on the 3D U-Net architecture, which has demonstrated strong performance in medical image segmentation tasks due to its encoder-decoder structure with skip connections<ref:2503.16988#pg5> The Vessel Lumen Structure Optimization Module (VLSOM) enhances the traditional 3D U-Net architecture by extracting vessel centerlines (Cl) and applying weight-adjusted supervision for vessel lumens based on their proximity to the centerlines<ref:2503.16988#pg3> The loss function design is defined as Loss = CE Loss + DICE Loss + 0.5 CLDICE Loss, where the weights are specified as Weight CE Loss = (wclass = 3, wcl = 15), Weight DICE Loss = (wclass = 3, wcl = 15, in lung mask), and Weight CL DICE Loss = (wclass = 3, wcl = 15, in lung mask)<ref:2503.16988#pg10> Performance Comparison Compared with the State of the Art (SOTA) method nnUNet-v2, our model demonstrates superior performance in pulmonary vessel segmentation for both CTPA and NCCT data<ref:2503.16988#pg5> The results show that our experimental model achieved Cl-DICE, Cl-Recall and Recall values of 0.892, 0.861, 0.924 for CTPA and 0.925, 0.903, 0.949 for NCCT data<ref:2503.16988#pg2> These metrics indicate that the model can segment more complete and intricate vascular tree branches compared to nnUNet-v2<ref:2503.16988#pg2> Furthermore, the method can segment more complete pulmonary vessel segmentation results when compared with public datasets like Parse2022 and HiPaS model<ref:2503.16988#pg8> Clinical Validation Four objective indicators were used to conduct the clinical visual evaluation: accuracy and robustness, branch abundance, assistance for diagnosis, and vascular continuity<ref:2503.16988#pg8> The average score for accuracy and robustness was 4.26 while the full score is 5<ref:2503.16988#pg8> The consistency of clinical evaluation by two radiologists showed good consistency across the four indicators, with a Kappa value of 0.916, 0.744, 0.674 and 0.796 respectively<ref:2503.16988#pg9> Future Directions The remaining challenges for pulmonary vessel segmentation model include insufficient segmentation accuracy for extreme small vessels (those with diameters less than 0.8 mm) and partial errors in artery-vein classification, particularly in complex or low-contrast regions<ref:2503.16988#pg10> Future improvements could focus on incorporating advanced techniques such as multi-scale feature extraction or attention mechanisms to better detect small vessels<ref:2503.16988#pg10> Additionally, expanding the training dataset and integrating anatomical priors could help reduce classification errors and improve robustness<ref:2503.16988#pg10> Further clinical studies should research the clinical values for the quantitative parameters derived from the segmented vessels such as vessel length, vessel volume, and tortuosity<ref:2503.16988#pg10> Pruning redundant reasoning tokens The study introduced a 3D UNet segmentation network with VLSOM module to enhance pulmonary vessel segmentation performance<ref:2503.16988#pg2> By leveraging combined learning from 213 CTPA and 214 NCCT data, the proposed method generated a single model which achieved high-accuracy and high-integrity automatic vascular segmentation on both 27 CTPA and 29 non-contrast CT images with Cl-Recall, ClDICE and Recall values of 0.879, 0.909, 0.934 for CTPA and 0.928, 0.936, 0.955 for NCCT respectively<ref:2503.16988#pg2> Furthermore, we proposed four visual assessment metrics for subjective scoring in clinical evaluation and the model demonstrated good performance on 21 independent external evaluation CT images from 21 patients with six different lung diseases (PV, PE, ILD, COPD, SLE and GGO), indicating potential for clinical pulmonary vessel analysis and disease diagnosis<ref:2503.16988#pg2> How the training signal is built The Vessel Lumen Structure Optimization Module (VLSOM) extracts the centerlines of the ground truth (GT) vessels and incorporates them as additional supervisory signals during training<ref:2503.16988#pg3> The three main modules within VLSOM include: 1.

Rosa: Vessel Centerline Extraction Outline, which processes GT vascular masks to extract centerlines as a structural cue<ref:2503.16988#pg5> 2.

Dev: Weighted Supervision, which applies different weights to the vessel regions based on their proximity to the centerlines<ref:2503.16988#pg5> 3.

Rosa: Loss function design, which incorporates CL-DICE Loss based on the vessel contour structure<ref:2503.16988#pg10> Where the gains come from The gains in performance stem from the structural cues provided by the VLSOM, specifically through extracting centerlines and applying weighted supervision<ref:2503.16988#pg3> The post-processing algorithm further refines results by performing Connected Component Analysis, removing outliers, selecting largest components, and iteratively refining connections between small disconnected components<ref:2503.16988#pg8> This iterative refinement process ensures that the overall vascular tree morphology is highly intact<ref:2503.16988#pg8> Pruning redundant reasoning tokens The study introduced a 3D UNet segmentation network with VLSOM module to enhance pulmonary vessel segmentation performance<ref:2503.16988#pg2> By leveraging combined learning from 213 CTPA and 214 NCCT data, the proposed method generated a single model which achieved high-accuracy and high-integrity automatic vascular segmentation on both 27 CTPA and 29 non-contrast CT images with Cl-Recall, ClDICE and Recall values of 0.879, 0.909, 0.934 for CTPA and 0.928, 0.936, 0.955 for NCCT data<ref:2503.16988#pg2> Furthermore, we proposed four visual assessment metrics for subjective scoring in clinical evaluation and the model demonstrated good performance on 21 independent external evaluation CT images from 21 patients with six different lung diseases (PV, PE, ILD, COPD, SLE and GGO), indicating potential for clinical pulmonary vessel analysis and disease diagnosis<ref:2503.16988#pg2> Where the gains come from The gains in performance stem from the structural cues provided by the VLSOM, specifically through extracting centerlines and applying weighted supervision<ref:2503.16988#pg3> The post-processing algorithm further refines results by performing Connected Component Analysis, removing outliers, selecting largest components, and iteratively refining connections between small disconnected components<ref:2503.16988#pg8> This iterative refinement process ensures that the overall vascular tree morphology is highly intact<ref:2503.

Paper summary: Rosa: So, to recap, this paper proposes a three dee image segmentation algorithm specifically built for automating pulmonary vessel segmentation using both CTPA and NCCT images > <ref:2503.16988#pg1,a 3D image segmentation algorithm>

Dev: The core idea is the Vessel Lumen Structure Optimization Module, or VLSOM, which uses extracted vessel centerlines to guide the segmentation process through weighted supervision and a specialized loss function called CL-DICE Loss >

Taro: They trained this model on a dataset that included both contrast-enhanced CTPA cases and non-contrast NCCT cases, with careful annotation done by two experienced radiologists to ensure high quality ground truth data >

Rosa: They achieved solid performance metrics across both modalities, hitting recall values up to zero point nine five for the NCCT data when compared against existing methods like nnUNet-v2 > <ref:2503.16988#pg1>

Dev: The dataset they used was quite balanced, with two hundred thirteen cases of CTPA and two hundred fourteen cases of NCCT, which allowed them to develop this versatile model that supports both imaging types > <ref:2503.16988#pg3,213 cases of CTPA and 214 cases of NCCT>

Taro: But the paper also lays out some limitations you have to consider. They flag that there's still a challenge with insufficient segmentation accuracy for very small vessels, those with diameters less than zero point eight millimeters >

Rosa: That means even if the model is good overall, it might struggle when trying to precisely capture those extremely fine vascular structures >

Dev: And another limitation they mentioned is partial errors in classifying arteries versus veins, especially in complex or low-contrast regions within the scans >

Conclusion: Rosa: So looking at the title, "High Accuracy Pulmonary Vessel Segmentation for Contrast and Non-contrast CT Images and Clinical Evaluation," it really captures the breadth of what they did here >

Dev: It means they successfully created a way to achieve high accuracy in segmenting pulmonary vessels regardless of whether the scan uses contrast or not, which is important for clinical work where you might face both scenarios >

Taro: The implication is that this unified approach could help speed up diagnosis and treatment planning because doctors would have a tool that works on more types of imaging data than before >

Rosa: It’s about moving beyond just focusing on one modality to create a system that is robust enough for real-world clinical use across different patient situations >

Dev: And the authors demonstrated this by showing their model performs well on both CTPA and NCCT data, which addresses the gap in research where NCCT segmentation has been less studied >

More episodes

← Home