In a recent study published in the journal Scientific Reports, researchers from the Republic of Korea introduced an automated workflow for measuring the size and shape of abdominal aortic aneurysm (AAA) using computed tomography (CT) images and deep learning techniques. They developed a semantic segmentation algorithm with active learning (AL) to accurately delineate the aorta, thrombus, calcification, and vessels from the CT scans.
DSC and HD95 of each class in stage 5 obtained using various networks. DSC and HD95 of (A,E) aorta, (B,F) thrombus, (C,G) calcification, and (D,H) vessels. Paired t-tests between stage 5 and other stages; *p < 0.05, **p < 0.005, ***p < 0.0005; DSC dice similarity coefficient, HD95 95% Hausdorff distance, SwinUNETR shifted-windows UNEt transformers. Image Credit: https://www.nature.com/articles/s41598-024-59735-8
Additionally, the researchers employed a computer-aided design (CAD) application programming interface (API) to automatically measure various critical landmarks of abdominal aortic aneurysms (AAA). This technology is essential for selecting suitable stent grafts and preventing complications associated with endovascular aneurysm repair (EVAR).
Background
AAA is a condition characterized by the enlargement and weakening of the abdominal aorta, the main blood vessel that supplies blood to the lower body. If left untreated, AAA can rupture, leading to life-threatening bleeding. To address this, minimally invasive procedures like EVAR are utilized, involving the insertion of a stent graft into the aneurysm to prevent further expansion and rupture. However, precise measurement of AAA size and shape is crucial for EVAR planning to select suitable stent grafts and avoid complications such as endoleaks, where blood leaks into the aneurysm sac post-operatively.
Current methods rely on medical imaging, particularly CT scans, for AAA evaluation before EVAR. However, manual segmentation and annotation of CT images by a human expert can be time-consuming and prone to inaccuracies, leading to poor inter-observer reproducibility. Therefore, there is a pressing need for automated methods to segment and measure AAA anatomy efficiently from CT images.
About the Research
In this paper, the authors developed an automated workflow comprising semantic segmentation and automated measurement. Semantic segmentation involves assigning labels to each pixel in an image, representing structures like the aorta, thrombus, calcification, or vessels. Automated measurement encompasses extracting key anatomical information from the segmented images, such as diameters, lengths, angles, and tortuosities of the aorta and iliac arteries.
The researchers used 300 CT scans of patients diagnosed with AAA from a single institution. They manually segmented the aorta, thrombus, calcification, and vessels from the CT images and used them as ground truths for training and testing various deep learning models, including UNEt Transformers (UNETR), shifted-windows UNEt Transformers (SwinUNETR), and no-new-U-Net (nnU-Net).
Moreover, they employed AL techniques to reduce the need for manual annotation by iteratively selecting the most informative data for labeling and training. Furthermore, the model performance was evaluated using metrics like the dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95) on 30 test scans.
For the automated measurement process, the authors employed the CAD software API to construct three-dimensional (3D) models of AAA from segmented images. Subsequently, they devised two algorithms to quantify seven pivotal AAA landmarks crucial for EVAR planning: aortic neck diameter, aneurysm diameter, diameters of the right and left iliac arteries, aortic neck length, common iliac artery tortuosity, and aortic neck angulation.
To validate these automated measurements, they conducted a comparison against conventional manual measurements based on two-dimensional (2D) images by medical professionals/doctors, employing Bland-Altman analysis.
Research Findings
The outcomes showed that the 3D U-Net model achieved the highest average DSC of 0.722, and the SwinUNETR model achieved the lowest average HD95 of 10.23 mm among the tested models. It also highlighted that AL improved the performance of the models as the stages progressed, especially for the thrombus and vessel classes. The segmentation time was significantly reduced from 13.74 min for manual segmentation to 2.09 min for AL-corrected segmentation using the 3D U-Net model.
Moreover, the study highlighted that the automated measurements exhibited consistency and accuracy when compared with manual measurements, showcasing mean differences ranging from 1.71 to 0.15 mm for diameters, lengths, and tortuosities, alongside negligible human error in the centerline division.
Unlike the conventional approach, the automated measurement method was less prone to human error and variation than the conventional method, as it used 3D models and centerlines instead of 2D images and cross-sectional planes. Additionally, the automated method reduced the time and labor required for manual measurement processes.
Furthermore, the authors demonstrated the feasibility and efficiency of their automated workflow for AAA segmentation and measurement using CT images and deep learning techniques. Their method can potentially improve the accuracy and reproducibility of the AAA anatomy evaluation and facilitate the EVAR planning and execution. Their method can also be applied to other medical image segmentation and measurement tasks that require high precision and robustness.
Conclusion
The paper summarized that the newly introduced workflow was valuable for AAA diagnosis and treatment planning. Moving forward, the researchers proposed further enhancements to their method, including validation on larger and more diverse datasets from multiple institutions, optimization of hyperparameters and model architectures, and the development of more advanced algorithms to reduce manual intervention.
Journal Reference
Kim, T., On, S., Gwon, J.G. et al. Computed tomography-based automated measurement of abdominal aortic aneurysm using semantic segmentation with active learning. Sci Rep 14, 8924 (2024). https://doi.org/10.1038/s41598-024-59735-8, https://www.nature.com/articles/s41598-024-59735-8.
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.