Related Experiment Video
Updated: May 5, 2026
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Carotid artery segmentation in computed tomography angiography (CTA) using multi-scale deep supervision with
Haodong Xie1, Hongmei Gu2, Minda Li2
1Department of Medical Informatics, Medical School of Nantong University, Nantong, China.
Insights
This study introduces an automated deep learning method for 3D carotid artery segmentation from CTA images, significantly improving accuracy and efficiency for diagnosing carotid artery disease (CAD). The novel Multi-Flux-Swin-Deepsup-UNet model offers a robust solution for clinical applications.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Cardiovascular Disease Research
Background:
- Carotid artery disease (CAD) is a significant risk factor for stroke, often diagnosed using time-consuming manual segmentation of computed tomography angiography (CTA) images.
- Current diagnostic methods for CAD lack automation, necessitating advanced techniques for efficient and accurate image analysis.
- The need for precise 3D carotid artery segmentation is critical for timely diagnosis and treatment planning.
Purpose of the Study:
- To develop and validate an automated deep learning (DL) model for accurate 3D carotid artery segmentation from CTA images.
- To enhance the efficiency and precision of CAD diagnosis through advanced image analysis techniques.
- To compare the performance of the proposed DL model against existing state-of-the-art methods.
Main Methods:
- Collected and annotated 214 CTA images from multiple hospitals.
- Employed a novel window/level adjustment for image preprocessing and augmentation.
- Utilized the Multi-Flux-Swin-Deepsup-UNet (MFSD-UNet) model, incorporating multi-scale deep supervision and multi-flux fusion for segmentation.
Main Results:
- The MFSD-UNet model achieved a high average Dice coefficient of 0.9119 and accuracy of 0.9819.
- Outperformed two state-of-the-art models with significantly better Dice coefficients (P<0.05).
- Ablation studies confirmed the superiority of the integrated Swin transformer and deep supervision components, demonstrating model robustness through seven-fold cross-validation.
Conclusions:
- A novel DL-based method for automatic 3D carotid artery segmentation from CTA images was successfully developed.
- The integration of Swin transformers, deep supervision, and data augmentation significantly improved segmentation accuracy and robustness.
- This automated approach provides valuable clinical support for CAD diagnosis and treatment, with potential for broader medical image segmentation applications.
Background:
Carotid artery disease (CAD) is a serious disease caused by atherosclerosis, resulting in reduced cerebral blood flow and an increased risk of stroke. Traditionally, CAD diagnosis involves manual segmentation of computed tomography angiography (CTA) images, a time-consuming and complex process. This study aimed to address the need for an automated and accurate method for three-dimensional (3D) carotid artery segmentation using deep learning (DL) techniques.
Methods:
A total of 214 CTA images from patients at the Affiliated Hospital of Nantong University and Nantong First People's Hospital were collected. The data were annotated using 3Dslicer software and calibrated by experienced radiologists. Preprocessing and augmentation of the CTA images were conducted using a novel window/level (W/L) adjustment method to enhance vascular imaging. The segmentation is performed using the Multi-Flux-Swin-Deepsup-UNet (MFSD-UNet) model, which incorporates multi-scale deep supervision and multi-flux fusion architecture. Performance was evaluated based on accuracy, dice coefficient, sensitivity, and specificity, and compared with state-of-the-art models. Ablation studies were conducted, removing the Swin transformer and deep supervision components to demonstrate the superiority of our method.
Results:
The proposed model showed excellent performance, achieving an average dice coefficient of 0.9119 and an accuracy of 0.9819, outperforming the average dice coefficients of 0.8770 and 0.8910 for the two state-of-the-art models. Furthermore, it demonstrated high stability across various segmentation categories. Ablation studies revealed that removing the Swin transformer and deep supervision components resulted in a decrease in the dice coefficient to 0.8630 and 0.8371. Significant differences were observed when comparing these four models with MFSD-UNet (P<0.05), and seven-fold cross-validations were performed on MFSD-UNet to demonstrate its robustness.
Conclusions:
This study introduced a novel DL-based method for automatic 3D carotid artery segmentation from CTA images. The integration of Swin transformers, deep supervision mechanisms, and innovative data augmentation techniques significantly enhanced the accuracy and robustness of segmentation. This method offers valuable support for the clinical diagnosis and treatment of CAD and exhibits great potential for future medical image segmentation.
