基于nnUNet多模型集成的超声心动图四腔室分割OA

Objective:To propose a multi-model ensemble framework based on no-new-Net(MME-nnUNet)on the basis of the problems of low resolution,noise interference,and insufficient annotation in the precise segmentation of echocardiograms for cardiac four-chamber,so as to improve the accuracy and robustness of segmentation.Methods:A total of 293 videos of apical four-chamber view in public dataset of cardiac ultrasound(CardiacUDC),which was issued in 2023 by Guangdong Provincial People's Hospital,were selected.The quality of data was optimized through multi-stage preprocessing(manual screening,morphological operation).Using the Residual U-shaped network(ResUNet)as the baseline model to construct a two dimension(2D)nnUNet model to extract features of single-frame images,and to generate pseudo-labels to relieve the issue of insufficient annotations of three dimension(3D)data.A 3D nnUNet model was designed to capture spatiotemporal correlations among consecutive frames.The optimized segmentation results of post-processing were preserved through integrated the outputs of 2D and 3D multi-models,and adopted the largest connected area.Results:In the test set,the Dice similarity coefficient(DSC)of MME-nnUNet was 0.946 6,and the average surface distance(ASD)was 0.435 2 mm,and the 95%Hausdorff distance(HD95)was 3.959 6 mm,which increased 2.89%,and decreased 0.5214 mm and 3.2794 mm than the baseline ResUNet model.Conclusion:The enhancement of DSC,and the decreases of ASD and HD95 through integrates the advantages of 2D and 3D models and through semi-supervised data augmentation and optimization of dynamic post-processing demonstrate that MME-nnUNet can enhance the accuracy of the segmentation for four cardiac chambers,which can provide reliably technical support for cardiac function assessment,and diagnosis and treatment for disease.