| 冉诗懿,鲁蓉,李沐宸,屈灿.超声深度学习模型用于诊断膀胱膨出并进行分型[J].中国医学影像技术,2025,41(10):1710~1714 |
| 超声深度学习模型用于诊断膀胱膨出并进行分型 |
| Ultrasound deep learning model for diagnosis and classification of cystocele |
| 投稿时间：2025-06-26 修订日期：2025-10-04 |
| DOI：10.13929/j.issn.1003-3289.2025.10.023 |
| 中文关键词: 膀胱膨出 超声检查 人工智能 |
| 英文关键词:cystocele ultrasonography artificial intelligence |
| 基金项目:湖南省自然科学基金(2023JJ30920)。 |
|
| 摘要点击次数: 281 |
| 全文下载次数: 39 |
| 中文摘要: |
| 目的 观察超声深度学习(DL)模型用于诊断膀胱膨出并进行分型的价值。方法 回顾性分析696例接受盆底超声检查的女性患者并划分模型开发数据集(n=576)与测试集(n=120),前者包括432例膀胱膨出、144例非膀胱膨出,后者包括90例膀胱膨出、30例非膀胱膨出。按8 ∶ 2比例随机将模型开发数据集分为训练集(n=460,含345例膀胱膨出及115例非膀胱膨出)与验证集(n=116,含87例膀胱膨出及29例非膀胱膨出)。基于训练集及验证集盆底超声数据采用Vision Transformer架构进行训练并建立DL模型以诊断膀胱膨出并判断其类型(无及Green Ⅰ、Ⅱ、Ⅲ型)。以超声科高级职称医师诊断结果为标准,评估模型诊断效能,并与2名超声科初级职称医师的诊断效能及诊断效率进行比较。结果 DL模型诊断验证集膀胱膨出并进行分型的宏平均精确率、F1分数、曲线下面积(AUC)及整体准确率分别为90.84%、89.28%、0.97及89.66%,在测试集分别为80.85%、79.92%、0.92及80.00%。2名超声科初级医师诊断测试集膀胱膨出并进行分型的整体准确率分别为70.00%(84/120)和68.33%(82/120),均低于DL模型(P=0.023、0.011)。DL模型每例诊断时间为0.098 s,人工每例诊断时间为46(36,56)s,前者诊断效率更高(P＜0.001)。结论 超声DL模型可用于诊断膀胱膨出并进行分型。 |
| 英文摘要: |
| Objective To explore the value of ultrasound deep learning (DL) model for diagnosis and classification of cystocele. Methods Totally 696 female patients who underwent pelvic floor ultrasound were retrospectively collected and divided into model development dataset (n=576) and test set (n=120). The former included 432 cases of cystocele and 144 cases of non-cystocele, while the latter included 90 cases of cystocele and 30 cases of non-cystocele. Patients in model development dataset were randomly divided into training set (n=460, including 345 cases of cystocele and 115 cases of non-cystocele) and validation set (n=116, including 87 cases of cystocele and 29 cases of non-cystocele) at the ratio of 8 ∶ 2. DL model was trained and established using Vision Transformer architecture based on pelvic floor ultrasound data in training and validation sets for diagnosis and classification of cystocele (non- or Green Ⅰ, Ⅱand Ⅲ type). Taken diagnostic results of senior ultrasound physicians as standard, the diagnostic efficacy of DL model was evaluated, and its diagnostic efficacy and efficiency were compared with those of 2 junior ultrasound physicians. Results The macro average precision, F1 score, area under the curve (AUC) and overall accuracy of DL model for diagnosis and classification of cystocele in validation set was 90.84%, 89.28%, 0.97 and 89.66%, respectively, while in test set was 80.85%, 79.92%, 0.92 and 80.00%, respectively. The overall diagnostic accuracy of 2 junior ultrasound physicians for diagnosis and classification of cystocele in test set was 70.00% (84/120) and 68.33% (82/120), respectively, both lower than that of DL model ( P=0.023, 0.011). The diagnostic time of DL model was 0.098 s for each case, of junior ultrasound physicians was 46 (36, 56) s for each case, the former had better diagnostic efficacy (P＜0.001). Conclusion Ultrasound DL model could be used for diagnosis and classification of cystocele. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
