Construction of an ultrasound dynamic image segmentation model for thyroid nodules
10.3760/cma.j.cn131148-20241213-00647
- VernacularTitle:甲状腺结节超声动态影像分割模型的构建
- Author:
Junpu HU
1
;
Jialu LI
;
Mengjie DOU
;
Gang WANG
;
Keyan LI
;
Xiaofang FU
;
Hao SUN
;
Changqin SUN
;
Duo SHI
;
Yan LIAO
;
Qiong WANG
;
Faqin LYU
Author Information
1. 锦州医科大学解放军总医院第三医学中心研究生培养基地，北京 100039
- Publication Type:Journal Article
- Keywords:
Thyroid nodule;
Dynamic ultrasound images;
Diffusion model
- From:
Chinese Journal of Ultrasonography
2025;34(6):518-524
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To construct a thyroid nodule segmentation model using ultrasound dynamic images and explore its potential for assisting in the screening of thyroid nodules.Methods:A total of 126 patients with thyroid nodules（comprising 150 nodules）who were diagnosed and treated at Xuzhou Cancer Hospital from April 2024 to December 2024 were prospectively enrolled. Two-dimensional ultrasound was performed to capture short-axis and long-axis video images of thyroid nodules，forming a dynamic ultrasound image dataset. The dataset was divided into training，validation，and test sets in a ratio of 6∶1∶3. After the training loss curve converged，the model that performed well on the validation set was selected for testing. Three-fold cross-validation was employed for training and testing. All 300 ultrasound videos were divided into three subsets. In each experiment，two subsets were used as the training set，and one subset was used as the test set to evaluate the model's generalization ability. A collaborative spatiotemporal diffusion model was established based on the dynamic trends and tissue texture details of thyroid nodules. Six widely used segmentation metrics were employed to evaluate the model's application capabilities.Results:The study included 126 patients with 150 thyroid nodules，300 dynamic ultrasound images，and video lengths of 3-4 seconds per nodule，resulting in 12 312 segmented images. The size of the thyroid nodules was（10.7 ± 10.6）mm（transverse diameter）×（8.4 ± 6.3）mm（anteroposterior diameter）. Among the nodules，62（41.3%）had clear boundaries，while 88（58.7%）had indistinct boundaries；61（40.7%）exhibited regular shapes，while 89（59.3%）were irregular；66（44.0%）had a taller-than-wide aspect ratio；and 70（46.7%）showed microcalcifications. The collaborative diffusion model based on dynamic ultrasound image segmentation achieved the following scores：a Jaccard score of（69.22 ± 0.03）%，a Dice score of（79.16 ± 0.18）%，a Precision score of（86.70 ± 0.17）%，a Recall score of（77.82 ± 0.04）%，an Sα score of（85.26 ± 0.01）%，and an Eθmn score of（90.58 ± 0.17）%. Compared to other models，this model demonstrated significant improvements across all evaluation metrics，achieving the highest values in each metric with increments of over 8% and 1%，respectively. Conclusions:The collaborative diffusion model with a dynamic controller，constructed based on dynamic ultrasound images of thyroid nodules，demonstrates excellent performance in ultrasound image segmentation. It improves the accuracy of thyroid nodule screening，thereby providing a valuable auxiliary diagnostic tool for clinical practice.
