A thyroid nodule ultrasound image segmentation method based on a feature-adaptive extraction and gated fusion mechanisms net
10.3760/cma.j.cn121382-20250628-00050
- VernacularTitle:基于特征自适应提取和门控融合机制网络的甲状腺结节超声图像分割方法
- Author:
Chengzi YAO
1
;
Han GUO
;
Zhiqiang JIA
Author Information
1. 咸阳市中心医院全科医学科，咸阳 712000
- Keywords:
Thyroid nodule;
Image segmentation;
Convolutional neural network;
Attention mechanism;
Medical image processing
- From:
International Journal of Biomedical Engineering
2025;48(4):357-364
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To propose a thyroid nodule ultrasound image segmentation method based on a feature-adaptive extraction and gated fusion mechanisms net (FAGF-Net).Methods:The FAGF-Net was constructed by designing a feature coupling encoder (FC-Encoder), which integrated a convolutional neural network and a Transformer to extract both local and global features. The perception of multi-scale geometric characteristics of nodules was enhanced by developing a multi-scale feature space coupling module (MFSC-Module), and a context-gated feature attention module (CGFA-Module) was introduced to filter redundant information and suppress noise interference. A retrospective analysis was conducted on 3 493 physician-annotated two-dimensional thyroid nodule ultrasound images obtained from the publicly available TN3K medical imaging dataset. The dataset was divided into training and validation sets at a ratio of 8∶2, resulting in 2 794 training images and 699 validation images. FAGF-Net was compared with several mainstream semantic segmentation models, including UNet, Deeplabv3, HRNet, PSPNet, and SegFormer using UNet as the baseline model. Additionally, module ablation experiments were performed to evaluate the effectiveness of each core module within the FAGF-Net. Visual comparisons were also conducted between the original ultrasound images, the annotated nodule regions, and the segmentation results produced by the different models.Results:The accuracy, intersection over union and Dice coefficient of the FAGF-Net method were 95.59%, 83.53% and 90.60%, respectively. These values were 3.77%, 2.29% and 3.55% higher than those of the UNet method (91.82%, 81.24% and 87.05%). Additionally, the FAGF-Net method achieved a frame rate of 19.8, meeting the requirements for image segmentation in medical environment scenarios. The module ablation experiments showed that compared with the UNet, the accuracy (92.85%), intersection over union (81.76%) and Dice coefficient (88.45%) of the FC-Encoder model were increased by 1.03%, 0.52% and 1.40%, respectively. The accuracy (93.13%), intersection over union (81.91%) and Dice coefficient (88.76%) of the CGFA-Module model were introduced separately, which were 1.31%, 0.67% and 1.71% higher than those of UNet, respectively. Compared with the FC-Encoder model alone, the accuracy (94.61%), intersection over union (82.45%) and Dice coefficient (89.54%) of the MFSC-Module increased by 1.76%, 0.69% and 1.09%, respectively when FC-Encoder was used as the feature double branch extraction encoder. Compared with the FC-Encoder model alone, the accuracy (94.39%), intersection over union (82.33%) and Dice coefficient (89.46%) of the CGFA-Module were increased by 1.54%, 0.57% and 1.01%, respectively, when FC-Encoder was used as the feature double branch extraction encoder. Visualization results showed that the FAGF-Net method produced smoother and more complete in boundary fitting and was more accurate at identifying small and low-contrast lesions.Conclusions:A thyroid nodule ultrasound image segmentation method based on FAGF-Net was proposed, which effectively improves the segmentation accuracy of thyroid nodule ultrasound image segmentation.
