基于U-net的心脏自动勾画模型的临床应用及改进

目的:拟分析基于不同医院数据的心脏自动勾画模型在临床应用中的适用性及其改进方法。方法:首先，建立基于U-net和Inception模块的心脏自动勾画网络。其次，收集不同治疗中心的患者数据:中国科学技术大学附属第一医院65例（数据1）、MICCAI2019比赛数据50例（数据2）、数据1和2的混合数据（数据3）、郑州大学第一附属医院50例（数据4）和郑州大学第一附属医院100例（数据5），分别训练得到模型1～5。然后，以郑州大学第一附属医院59例患者作为测试集，使用形状相似性系数（DSC）评估该测试集在不同模型上的分割精度，比较模型之间的差别。最后，将模型3作为心脏预训练模型，采用数据5进行模型再训练，分别测试3组实验（20例/次×5次、10例/次×10次、5例/次×20次）对心脏预训练模型的改进情况。结果:测试集在模型1~5中的平均DSC为0.926、0.932、0.939、0.941和0.950。在再训练过程中，模型在20例/次×5次的实验中表现更稳定。结论:基于不同医院的数据训练模型在心脏自动勾画的临床应用上表现存在差异，使用本地医院数据进行训练的模型预测精度更高。对于非本地数据训练的模型，基于本地数据再训练可以有效提高模型预测的精度，其中以20例/次的再训练方式效果较好。

Clinical application and improvement of U-net-based model for automatic segmentation of the heart

Abstract: Objective To investigate the clinical applicability of the model established based on the data from different hospitals for the automatic segmentation of the heart and to discuss the methods to improve the model. Methods A network based on U-net and Inception module was firstly constructed for the automatic segmentation of the heart, and the clinical data from different hospitals were collected, including 65 cases from the First Affiliated Hospital of University of Science and Technology of China (data 1), 50 cases from MICCAI2019 match data (data 2), the mixed data of data 1 and data 2 (data 3), 50 cases from the First Affiliated Hospital of Zhengzhou University (data 4) and 100 cases from the First Affiliated Hospital of Zhengzhou University (data 5). The collected data were trained for obtaining 5 different models. Then, with the clinical data of another 59 patients from the First Affiliated Hospital of Zhengzhou University as test set, the segmentation accuracies of test set on different models were evaluated using Dice similarity coefficient (DSC), and the differences in segmentation accuracies among different models were also compared. Finally, model 3 was used as a pre-trained model of the heart, and the model was retrained with data 5. Three groups of experiments (20 cases each time × 5 times, 10 cases each time× 10 times, 5 cases each time× 20 times) were carried out to observe the improvement of the pre-trained model. Results The average DSC of the test set based on models 1 to 5 was 0.926, 0.932, 0.939, 0.941 and 0.950, respectively. During the retraining of the pre-trained model of the heart, the model was more stable in the experiment of 20 cases each time × 5 times. Conclusion The trained model established based on the data from different hospitals has different performances in the automatic segmentation of the heart, and the model trained with local hospital data has a higher prediction accuracy. For the model based on non-local data training, the retraining with local data can effectively improve the accuracy of model prediction, in which the retraining with 20 cases each time has the optimal performance.