构建多病种心脏MRI跨设备智能分割算法

An intelligent cross-device segmentation algorithm for multi-disease cardiac MR images

目的构建多病种心脏磁共振成像(magnetic resonance imaging,MRI)跨设备智能分割算法,提升模型在多病种条件下、不同影像设备中的通用性。方法利用MICCAI 2020公开的M&Ms Challenge心脏磁共振数据集(n=320)作为研究对象,针对现有分割模型因小样本数据训练导致泛化能力差的问题,提出不平衡相似度优化损失函数USOLoss,改进主流的UNet、DeepLabV3+、TransUNet算法,对不同病种(疾病组)和不同影像设备(设备组)的心脏磁共振成像进行分割,并进行内外部数据验证。结果利用戴斯相似系数(Dice similarity cofficient,DSC)和豪斯多夫距离(Hausdorff distance,HD)评估模型的性能,其中疾病组模型最佳分割结果DSC为0.845(扩张型心肌病,n=20)、0.811(肥厚型心肌病,n=20)、0.833(健康受试者,n=20)和0.816(其他病种,n=0.62),HD为3.05、2.53、2.15和2.36 mm;设备组模型最佳分割结果DSC为0.830(飞利浦,n=20)、0.844(西门子、n=20)、0.843(佳能,n=20)和0.815(通用电气,n=50),HD指标分别为1.96、2.92、1.67和2.08 mm。与未使用构建算法的模型相比,使用USOLoss构建的模型各项测试结果均得到了提升(P<0.05)。结论不平衡相似度优化损失函数全面提升了现有主流深度学习UNet、DeepLabV3+和TransUNet网络模型性能,降低了不同疾病类型和不同影像设备对分割性能的影响。

Objective To build a multi-disease cardiac magnetic resonance imaging(MRI)cross-device intelligent segmentation algorithm in order to improve the generality of the model in multi-disease conditions and different imaging devices.MethodsBased on the M&Ms Challenge cardiac dataset(n=320)which was accepted by the international conference on Medical Image Computing and Computer Assisted Intervention(MICCAI)in 2020 as the research object,we proposed the hybrid loss function USO Loss to address the problem of poor generalization ability of existing segmentation models due to small sample data training,and improve the mainstream UNet,DeepLabV3+,and TransUNet algorithms in the multi-disease for different imaging device cardiac MRI data for segmentation.The experiments were divided into disease(n=320)and device groups(n=320),and internal and external data validation was performed.Results Dice similarity cofficient(DSC)and Hausdorff distance(HD)were used to evaluate the performance of model.The optimal segmentation results of DSC in disease group were 0.845(dilated cardiomyopathy,n=20),0.811(hypertrophic cardiomyopathy,n=20),0.833(health volunteers,n=20)and 0.816(others,n=62),and of HD were 3.05,2.53,2.15 and 2.36 mm.The optimal segmentation results of DSC in device group were 0.830(Philips,n=20),0.844(Siements,n=20),0.843(Canon,n=20)and 0.815(General Electronics,n=50),and of HD were 1.96,2.92,1.67 and 2.08 mm.Compared with the models unused construction algorithm,models used USOLoss had improved results in all tests(P<0.05).Conclusion The proposed USOLoss comprehensively improves the existing mainstream deep learning UNet,DeepLabV3+and TransUNet network models performance and reduces the impact of different diseases and imaging devices on the segmentation algorithm.