| 肖丽珊,赵毅哲,李昱臣,闫萌萌,杜梅霞,赵诚,刘满华,宁春平.改良ResNet18轻量深度学习模型基于第一跖趾关节声像图自动检测痛风性关节炎病变[J].中国医学影像技术,2025,41(5):783~787 |
| 改良ResNet18轻量深度学习模型基于第一跖趾关节声像图自动检测痛风性关节炎病变 |
| Improved ResNet18 lightweight deep learning models for automatically detecting gouty arthritis lesions based on ultrasonogram of the first metatarsophalangeal joint |
| 投稿时间：2024-08-15 修订日期：2025-03-24 |
| DOI：10.13929/j.issn.1003-3289.2025.05.019 |
| 中文关键词: 关节炎,痛风性 超声检查 深度学习 跖趾关节 |
| 英文关键词:arthritis, gouty ultrasonography deep learning metatarsophalangeal joint |
| 基金项目:国家重点研发计划项目(2022YFC2503305)、青岛大学附属医院"临床医学与X"科研项目(QDFY+X2024133)。 |
| 作者 | 单位 | E-mail | | 肖丽珊 | 青岛大学附属医院腹部超声科, 山东 青岛 266000 | | | 赵毅哲 | 上海交通大学电子信息与电气工程学院, 上海 201100
上海交通大学人工智能研究所人工智能关键实验室, 上海 201100 | | | 李昱臣 | 青岛大学附属医院腹部超声科, 山东 青岛 266000 | | | 闫萌萌 | 青岛大学附属医院腹部超声科, 山东 青岛 266000 | | | 杜梅霞 | 青岛大学附属医院腹部超声科, 山东 青岛 266000 | | | 赵诚 | 青岛大学附属医院腹部超声科, 山东 青岛 266000 | | | 刘满华 | 上海交通大学电子信息与电气工程学院, 上海 201100
上海交通大学人工智能研究所人工智能关键实验室, 上海 201100 | | | 宁春平 | 青岛大学附属医院腹部超声科, 山东 青岛 266000 | 152081340@qq.com |
|
| 摘要点击次数: 49 |
| 全文下载次数: 13 |
| 中文摘要: |
| 目的 观察改良ResNet18轻量深度学习(DL)模型基于第一跖趾关节(MTP1)声像图自动检测痛风性关节炎(GA)病变的价值。方法 回顾性纳入260例疑诊痛风患者的2 401幅MTP1声像图,按4∶[1划分训练集(209例1 910幅图像)与测试集(51例491幅图像),手动标注图中GA病变。经预处理后采用ResNet18轻量网络构建DL模型,用于识别超声图像类别为正常或异常(存在任何GA表现),采用5折交叉验证法评估以2个、3个、4个及6个残差块构建的模型1、2、3、4的效能,记录各模型计算量及参数量,于测试集进行验证,以筛选最佳DL模型。结果 模型1、2、3、4的计算量分别为7 558.27、2 963.73、4 012.33及6 093.39 M,参数量分别为4.61、4.91、4.91及5.28 M,模型2的计算量最少且参数量仅略多于模型1。测试集中4个模型的准确率及曲线下面积差异均无统计学意义(P均＞0.05);模型2的敏感度高于而特异度低于模型3(P均＜0.05),为最佳DL模型。结论 所获改良ResNet18轻量DL模型均可基于MTP1声像图自动检测GA;其中以模型2为最佳。 |
| 英文摘要: |
| Objective To explore the value of improved ResNet18 lightweight deep learning (DL) models for automatically detecting gouty arthritis (GA) based on ultrasonogram of the first metatarsophalangeal joint (MTP1). Methods A total of 2 401 ultrasonograms obtained from 260 patients with suspected gout who underwent MTP1 ultrasound examination were included and divided into training set (1 910 ultrasonograms from 209 cases ) and test set (491 ultrasonograms from 51 cases) at the ratio of 4 ∶ 1. GA lesions on ultrasonograms were manually labeled. After preprocessing, ResNet18 lightweight network was used to construct DL models for identifying the ultrasonogram category was normal or abnormal (with any manifestation of GA). Five-fold cross-validation method was adopted to evaluate the efficacy of the DL models constructed with 2, 3, 4 or 6 residual blocks, i.e. model 1, 2, 3 and 4,respectively, and the computational cost and the amount of parameters of each model were recorded. The efficacy of the models were verified using test set, and the best DL model was screened. Results The computational cost of model 1, 2, 3 and 4 was 7 558.27, 2 963.73, 4 012.33 and 6 093.39 M, respectively, while the amount of parameters was 4.61, 4.91, 4.91 and 5.28 M, respectively. Model 2 had the least computational cost with parameters only slightly more than model 1. In test set, no significant difference of accuracy nor the area under the curve was found among 4 models (all P＞0.05). The sensitivity of model 2 was higher than that of model 3, while its specificity was lower only than that of model 3 (both P＜0.05), hence model 2 was the best DL model. Conclusion Improved ResNet18 lightweight DL models could be used for automatically detecting GA based on ultrasonogram of MTP1, among which model 2 was the best one. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
