Recognition of thyroid nodules using transport-based contrast-enhanced ultrasound sequence metric learning
刘春蕊 1万鹏 2孔文韬 1张道强 2姚静 1刘新艳
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 1南京大学医学院附属鼓楼医院超声医学科，南京 210008
2. 2南京航空航天大学计算机科学与技术学院，南京 210016
折叠
摘要
目的 建立甲状腺超声造影血流灌注相似性度量学习方法，对临床常见的甲状腺结节进行分类诊断。 方法 收集2017年1月至2020年12月南京鼓楼医院145例甲状腺结节超声造影病例，经手术病理明确诊断，包括20例结节性甲状腺肿，33例甲状腺滤泡状腺瘤，68例甲状腺微小乳头状癌和24例甲状腺乳头状癌。引入时序约束最优传输构建甲状腺超声造影度量学习（TCSML）方法，自动对齐局部造影增强模式并量化为动态血流灌注差异，并与已有的12种度量学习或多视图学习方法比对并辅助甲状腺结节进行分型鉴别。 结果 在与已有的12种度量学习或多视图学习方法的对比中，该TCSML模型的平均诊断准确度75.14%，灵敏度68.85%，精准度78.85%和F1分数69.79%，高于其他学习方法（均P<0.05）。模型的最优输入序列长度为7，最佳时序正则系数为0.05，最优特征投影维度为40，最优结构化范数约束为0.004。 结论 该模型能够有效辅助临床医生初步判断甲状腺分型类别，并提供主要的局部增强模式，为超声医生提供直观地诊断参考。 Objective To establish a metric learning approach that measures the similarity in blood perfusion on contrast-enhanced ultrasound (CEUS) of the thyroid gland, and to validate its use in categorical diagnosis of common thyroid nodules in clinical settings. Methods A total of 145 patients with thyroid nodules who underwent CEUS in Nanjing Drum Tower Hospital between January 2017 and December 2020 were included. According to post-surgical pathology, these patients comprised 20 cases of nodular goiter, 33 of follicular thyroid adenoma, 68 of papillary thyroid microcarcinoma, and 24 of papillary thyroid carcinoma. A temporally regularized optimal transport algorithm was introduced to develop an approach of thyroid contrast-enhanced sonography metric learning (TCSML) that automatically aligned local patterns of contrast enhancement and quantified the dynamic difference in blood perfusion. Then we compared TCSML to other 12 available approaches of metric learning or multi-view learning for the diagnostic performance in classification of thyroid nodules. Results Compared with the other 12 available approaches of metric learning or multi-view learning, TCSML on average yielded a diagnostic accuracy of 75.14%, a sensitivity of 68.85%, an accuracy of 78.85% and an F1 score of 69.79%, outperforming its counterparts (all P<0.05). The TCSML model presented an optimal input sequence length of 7, optimal temporal regularization coefficient of 0. 05, an optimal feature projection dimension of 40, and an optimal structured norm constraint of 0.004. Conclusion The TCSML model can effectively help clinicians in preliminary classification of thyroid nodules and, with a highlight on major modes for local enhancement, provides an intuitive diagnostic reference for sonographists.
Abstract
Objective To establish a metric learning approach that measures the similarity in blood perfusion on contrast-enhanced ultrasound (CEUS) of the thyroid gland, and to validate its use in categorical diagnosis of common thyroid nodules in clinical settings. Methods A total of 145 patients with thyroid nodules who underwent CEUS in Nanjing Drum Tower Hospital between January 2017 and December 2020 were included. According to post-surgical pathology, these patients comprised 20 cases of nodular goiter, 33 of follicular thyroid adenoma, 68 of papillary thyroid microcarcinoma, and 24 of papillary thyroid carcinoma. A temporally regularized optimal transport algorithm was introduced to develop an approach of thyroid contrast-enhanced sonography metric learning (TCSML) that automatically aligned local patterns of contrast enhancement and quantified the dynamic difference in blood perfusion. Then we compared TCSML to other 12 available approaches of metric learning or multi-view learning for the diagnostic performance in classification of thyroid nodules. Results Compared with the other 12 available approaches of metric learning or multi-view learning, TCSML on average yielded a diagnostic accuracy of 75.14%, a sensitivity of 68.85%, an accuracy of 78.85% and an F1 score of 69.79%, outperforming its counterparts (all P<0.05). The TCSML model presented an optimal input sequence length of 7, optimal temporal regularization coefficient of 0. 05, an optimal feature projection dimension of 40, and an optimal structured norm constraint of 0.004. Conclusion The TCSML model can effectively help clinicians in preliminary classification of thyroid nodules and, with a highlight on major modes for local enhancement, provides an intuitive diagnostic reference for sonographists.
万方数据