探究联合深度学习和三尖瓣彩色多普勒超声血流频谱图对心脏三尖瓣反流严重程度进行智能评估的可行性和准确性。

选取2022年10月至2023年4月在四川省人民医院进行超声心动图检查诊断三尖瓣反流的患者2629例，获取三尖瓣彩色多普勒超声血流频谱图2629张。由2位高年资专科医师使用LabelMe软件，把每一个完整周期标注出来，按照9∶1的比例划分为训练集和验证集，采用YOLOv5深度学习网络实现反流心动周期的自动标注。采用平均精度、精确度、召回率对心动周期自动标注模型的性能进行评价。由2位高年资专科医师对单个心动周期反流频谱图的反流严重程度进行轻、中、重度划分，制作成数据集，按照8∶1∶1的比例划分为训练集、验证集和测试集，采用ConvNeXt深度学习网络实现反流严重程度的评估。采用精确率、召回率、特异性和F1分数对反流程度评估模型的诊断性能进行评价。

模型智能检测反流心动周期的平均精度为0.979，精确度为0.951，召回率为0.972；模型智能评估反流严重程度的预测结果的加权平均精确率为0.938，召回率为0.928，特异性为0.964，F1分数为0.953，均处于较高值范围。

基于深度学习的超声心动图三尖瓣反流严重程度智能评估方法的评估结果与超声医师人工评估的结果一致性较高，具有较好的可行性和诊断准确性。

To explore the feasibility and accuracy of deep learning-based intelligent assessment of the severity of tricuspid regurgitation by color Doppler echocardiographic flow spectrometry of the tricuspid valve.

A total of 2629 patients with tricuspid regurgitation diagnosed by echocardiography at Sichuan Provincial People's Hospital from October 2022 to April 2023 were selected, and 2629 color Doppler echocardiographic flow spectrometry images of the tricuspid valve were obtained. Two senior specialists used LabelMe software to label each complete cycle, and divided them into either a training set or a validation set at a ratio of 9:1. The YOLOv5 deep learning network was used to automatically annotate the regurgitation cardiac cycle. The performance of the automatic annotation model for the cardiac cycle was evaluated using average precision, accuracy, and recall. Two senior specialists classified the severity of regurgitation into mild, moderate, and severe levels for each single cardiac cycle regurgitation spectrometry image, and created datasets. The datasets were divided into a training set, a validation set, and a test set at a ratio of 8:1:1. The ConvNeXt deep learning network was used to evaluate the severity of regurgitation. The diagnostic performance of the regurgitation severity assessment model was evaluated using accuracy, recall, specificity, and F1 score.

The model's average precision for intelligent detection of regurgitation heart cycle was 0.979, with an accuracy of 0.951 and recall rate of 0.972. The weighted average precision of the prediction results of the model's intelligent assessment of the severity of regurgitation was 0.938, with a recall rate of 0.928, specificity of 0.964, and F1 score of 0.953, all within a high value range.

The evaluation results of the intelligent assessment method for the severity of tricuspid regurgitation in echocardiography based on deep learning are highly consistent with those manually assessed by ultrasonographers, demonstrating good feasibility and diagnostic accuracy.