METHODS: We trained convolutional neural networks (UNet/Attention UNet) on 1171 images from 196 recordings (54 patients) to detect and segment flow convergence zones in 2-D color-Doppler images. Different preprocessing schemes and model architectures were compared. Flow convergence surface areas were estimated, accounting for non-hemispherical convergence, and regurgitant volume (RVol) was computed by integrating the flow rate over time. EasyPISA was retrospectively applied to 26 MR patient examinations, comparing results with reference PISA RVol measurements, severity grades, and cMRI RVol measurements for 13 patients.
RESULTS: The UNet trained on duplex images achieved the best results (precision: 0.63, recall: 0.95, dice: 0.58, flow rate error: 10.4 ml/s). Mitigation of false-positive segmentation on the atrial side of the mitral valve was achieved through integration with a mitral valve segmentation network. The intraclass correlation coefficient was 0.83 between EasyPISA and PISA, and 0.66 between EasyPISA and cMRI. Relative standard deviations were 46% and 53%, respectively. Receiver operator characteristics demonstrated a mean area under the curve between 0.90 and 0.97 for EasyPISA RVol estimates and reference severity grades.
CONCLUSIONS: EasyPISA demonstrates promising results for fully automated integrated PISA measurements in MR, offering potential benefits in workload reduction and mitigating inter-observer variability in MR assessment.
方法:我们对来自196个记录(54名患者)的1171个图像进行了卷积神经网络(UNet/AttentionUNet)的训练,以检测和分割二维彩色多普勒图像中的血流会聚区。比较了不同的预处理方案和模型架构。估计了流动会聚表面积,考虑到非半球形收敛,和反流体积(RVol)通过随时间积分流速来计算。EasyPISA应用于26例MR患者检查,将结果与参考PISARVol测量结果进行比较,严重性等级,和13例患者的cMRIRVol测量。
结果:在双工图像上训练的UNet取得了最好的结果(精度:0.63,召回率:0.95,骰子:0.58,流速误差:10.4ml/s)。通过与二尖瓣分割网络集成,可以减轻二尖瓣心房侧的假阳性分割。EasyPISA和PISA之间的组内相关系数为0.83,EasyPISA和cMRI之间为0.66。相对标准偏差分别为46%和53%,分别。接收器操作员特征表明,EasyPISARVol估计值和参考严重程度等级的曲线下平均面积介于0.90和0.97之间。
结论:EasyPISA证明了在MR中全自动集成PISA测量的有希望的结果,在减少MR评估中的工作量和减轻观察者之间的差异方面提供潜在的好处。