Ultrasound radiomics combined with machine learning for early diagnosis of seronegative hashimoto’s thyroiditis
10.3760/cma.j.cn115807-20241228-00396
- VernacularTitle:基于超声影像组学结合机器学习早期诊断甲状腺肿瘤合并抗体阴性桥本甲状腺炎
- Author:
Wenjun WU
1
;
Chang LIU
;
Shengsheng YAO
;
Daming LIU
;
Yuan LUO
;
Yihan SUN
;
Ting RUAN
;
Mengyou LIU
;
Li SHI
;
Mingming XIAO
;
Qi ZHANG
;
Zhengshuai LIU
;
Xingai JU
;
Jiahao WANG
;
Xiang FEI
;
Li LU
;
Yang GAO
;
Ying ZHANG
;
Liying GONG
;
Xuanyu CHEN
;
Wanli ZHENG
;
Xiali NIU
;
Xiao YANG
;
Huimei CAO
;
Shijie CHANG
;
Zuoxin MA
;
Jianchun CUI
Author Information
1. 辽宁中医药大学辽宁省人民医院研究生培养基地，沈阳 110013
- Publication Type:Journal Article
- Keywords:
Radiomics;
Hashimoto’s thyroiditis;
Serum-negative Hashimoto’s thyroiditis;
Machine learning;
Ultrasound imaging
- From:
Chinese Journal of Endocrine Surgery
2025;19(3):313-319
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To evaluate the value of ultrasound radiomics combined with machine learning for early diagnosis of seronegative Hashimoto’s thyroiditis (SN-HT) .Methods:This retrospective study included 164 patients from Liaoning Provincial People’s Hospital , Lixin County People’s Hospital, Linghai Dalinghe Hospital, Fengcheng Phoenix Hospital, who underwent thyroidectomy for solitary nodules with normal thyroid function between Nov. 2016 and Jan. 2024. Postoperative pathology confirmed Hashimoto’s thyroiditis (HT) in some cases, who were further categorized into antibody-positive and antibody-negative groups based on serum antibody status. Patients without Hashimoto’s thyroiditis served as the control group. A total of 298 ultrasound images were analyzed. Radiomics features were extracted from hypoechoic non-nodular areas within 0.5 cm surrounding the tumor. Two senior pathologists and two senior ultrasound physicians independently assessed lymphocytic infiltration, eosinophilic changes of follicular epithelium, and the proportion of hypoechoic areas in pathology and ultrasound images, respectively. A machine learning model, CCH-NET, was developed using linear regression and t-distributed stochastic neighbor embedding (t-SNE) techniques. The dataset was divided into a training set (80%) and a validation set (20%) to compare the diagnostic accuracy of CCH-NET with that of senior ultrasound physicians. Results:In internal validation, CCH-NET achieved a diagnostic accuracy of 88.89% for both antibody-positive and antibody-negative groups, significantly higher than the 66.67% accuracy of senior ultrasound physicians ( P<0.01). In external validation, CCH-NET achieved 75.00% and 66.67% accuracy for the two groups, compared to 50.00% by senior ultrasound physicians. For the control group, both methods achieved 93.33% accuracy. The AUC of CCH-NET was 0.848, outperforming senior ultrasound physicians (0.681) ,demonstrating superior diagnostic performance. Conclusion:The radiomics-based CCH-NET model, using non-nodular hypoechoic areas as a specific indicator, can accurately identify early SN-HT in euthyroid patients. It significantly outperforms senior ultrasound physicians, improving diagnostic accuracy and reducing missed diagnoses.
