The 3D ultrasound features of patients with pelvic organ prolapses (POP) were analyzed using deep learning algorithms to explore the diagnostic efficacy of deep learning in identifying the types of POP and evaluate the feasibility of pelvic floor 3D ultrasound in diagnosing POP. Forty-four POP patients were included as the research objects, including 17 cases of anterior cavity prolapse, 23 cases of middle cavity prolapse, and four cases of posterior cavity prolapse. Twenty healthy women without POP were the control group. The 3D perineum ultrasound images of the patients were collected. A feature classification model for pelvic floor ultrasound images was built using deep learning algorithms. The proposed model was then compared with three convolutional neural network (CNN) models: support vector machine (SVM), radial basis function (RBF), and K-nearest neighbor (KNN), to evaluate its diagnostic efficiency in identifying the types of POP. Furthermore, the POP-Q score was performed based on ultrasound features. The results showed that the area under the curve (AUC) of the deep learning model was 0.79, and the 95% confidence interval was 0.76–0.80, higher than those of the CNN models. The accuracy (0.86), sensitivity (0.89), and specificity (0.84) of the deep learning model were higher than those of the three CNN models. The pelvic mouths of the study group and the control group were significantly more extensive than that of the nonproductive group (p < 0.05). The levator ani muscle of the POP group was significantly more extensive than that of the control group, and the difference was statistically significant (p < 0.05). The model based on deep learning has high efficiency in identifying the types of POP. The 3D ultrasound examination has important clinical significance for POP diagnosis.

使用深度学习算法分析了盆腔器官脱垂（POP）患者的3D超声特征，以探索深度学习在识别POP类型方面的诊断功效，并评估盆底3D​​超声在POP诊断中的可行性。研究对象为44例POP患者，其中前腔脱垂17例，中腔脱垂23例，后腔脱垂4例。对照组中有20名没有POP的健康女性。收集患者的3D会阴超声图像。使用深度学习算法建立了骨盆底超声图像的特征分类模型。然后将提出的模型与三种卷积神经网络（CNN）模型进行比较：支持向量机（SVM），径向基函数（RBF），和K近邻（KNN），以评估其在识别POP类型方面的诊断效率。此外，根据超声特征进行POP-Q评分。结果表明，深度学习模型的曲线下面积（AUC）为0.79，95％置信区间为0.76-0.80，高于CNN模型。深度学习模型的准确性（0.86），敏感性（0.89）和特异性（0.84）高于三个CNN模型的准确性。研究组和对照组的骨盆口比非生产组的骨盆口明显更宽（结果表明，深度学习模型的曲线下面积（AUC）为0.79，95％置信区间为0.76-0.80，高于CNN模型。深度学习模型的准确性（0.86），敏感性（0.89）和特异性（0.84）高于三个CNN模型的准确性。研究组和对照组的骨盆口比非生产组的骨盆口明显更宽（结果表明，深度学习模型的曲线下面积（AUC）为0.79，95％置信区间为0.76-0.80，高于CNN模型。深度学习模型的准确性（0.86），敏感性（0.89）和特异性（0.84）高于三个CNN模型的准确性。研究组和对照组的骨盆口比非生产组的骨盆口明显更宽（p  ＜0.05）。POP组的肛提肌比对照组的广泛得多，差异具有统计学意义（p  <0.05）。基于深度学习的模型在识别POP类型方面具有很高的效率。3D超声检查对POP诊断具有重要的临床意义。