探讨基于全段尿道动态参数人工智能(AI)膀胱膨出自动识别模型，在盆底超声检查中对膀胱膨出的辅助诊断价值。

选择2024年7至12月中国医科大学附属盛京医院门诊被盆底超声诊断为膀胱膨出的103例与前盆腔结构正常者的97例，共计200例受试者为研究对象。将这200例受试者按照7∶3分别纳入训练集(n＝140)和测试集(n＝60)。训练集140例受试者中，膀胱膨出和前盆腔结构正常者分别为73、67例，测试集60例受试者中，膀胱膨出和前盆腔结构正常者均为30例。采集训练集受试者静息及最大Valsalva动作下的动态超声图像，手动勾画尿道轮廓，并提取13项尿道运动特征参数纳入训练集，构建基于支持向量机(SVM)及随机森林(RF)模型，采用5折交叉验证得到的5个权重，对测试集受试者进行测试，并分别评估SVM和RF模型，对膀胱膨出的诊断效能。本研究遵循的程序符合中国医科大学附属盛京医院医学伦理委员会要求，获得其批准(伦理审批号：2024PS1218K)，并且与所有受试者签署临床研究知情同意书。

①训练集和测试集受试者年龄、人体质量指数(BMI)、身高、体重、孕次、产次及民族、职业构成比等比较，差异均无统计学意义(P>0.05)。②测试集受试者中，SVM模型的准确率、精确度、召回率、F1分数及受试者工作特征的曲线下面积(ROC-AUC)分别为0.783、0.718、0.933、0.811、0.893，均优于RF模型的0.733、0.659、0.966、0.783、0.876。其中，SVM模型第1折权重文件表现最佳(ROC-AUC为0.904)。③混淆矩阵结果显示，SVM和RF模型对膀胱膨出的正确识别率分别为93.3%(28/30)，96.7%(29/30)；SVM和RF模型对前盆腔结构正常的正确识别率分别为63.3%(19/30)和50.0%(15/30)。

本研究基于全段尿道动态参数的AI自动识别模型，对受试者膀胱膨出具有良好的诊断效能，可为传统盆底超声检查提供客观的辅助决策工具。未来在膀胱膨出患者的早期筛查及术后随访中具有广泛应用前景，但是仍需扩大样本量的多中心临床试验进行验证，通过优化算法推动上述模型的临床转化。

To evaluate the auxiliary diagnostic value of an artificial intelligence (AI)-based automatic recognition model for cystocele in pelvic floor ultrasound examination.

A total of 200 female outpatients who underwent pelvic floor ultrasound examination at Shengjing Hospital of China Medical University between July and December 2024 were included. Among them, 103 cases were diagnosed with cystocele and 97 had normal anterior pelvic compartment structures. The subjects were randomly divided into a training set (n=140) and a testing set (n=60) in a 7∶3 ratio. In the training set, 73 had cystocele and 67 were normal; in the testing set, both groups included 30 cases. Dynamic pelvic floor ultrasound images at rest and during maximal Valsalva maneuver were collected. Urethral contours were manually delineated, and 13 urethral mobility parameters were extracted. Based on these features, support vector machine (SVM) and random forest (RF) models were trained using five-fold cross-validation, and the resulting five sets of weights were used to evaluate diagnostic performance in the testing set. The study was approved by the Ethics Committee of Shengjing Hospital, China Medical University (Approval No. 2024PS1218K), and all participants provided written informed consent.

①There were no statistically significant differences in age, body mass index (BMI), height, weight, gravidity, parity, ethnicity, or occupational composition between the training and testing sets (P>0.05). ②In the testing set, the SVM model demonstrated an accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (ROC-AUC) of 0.783, 0.718, 0.933, 0.811, and 0.893, respectively, all superior to those of the RF model (0.733, 0.659, 0.966, 0.783, and 0.876). Among the five SVM model, the first fold performed best with a ROC-AUC of 0.904. ③Confusion matrix analysis showed that the correct identification rates of cystocele by the SVM and RF models were 93.3% (28/30) and 96.7% (29/30), respectively; for normal anterior compartment structures, the rates were 63.3% (19/30) and 50.0% (15/30), respectively.

The AI-based automatic recognition model using full-length urethral dynamic parameters demonstrates favorable diagnostic performance in identifying cystocele and may serve as an objective auxiliary tool to conventional pelvic floor ultrasound. It holds promise for early screening, subtype classification, and postoperative follow-up of cystocele. However, further studies with larger sample sizes and multi-center validation are required to optimize the algorithm and facilitate clinical translation.