基于多任务学习的native T1 mapping图像对高血压心脏病和肥厚型心肌病的自动分类

Multi-task learning for automated classification of hypertensive heart disease and hypertrophic cardiomyopathy using native T1 mapping

目的基于多任务学习算法, 使用native T1 mapping图像对高血压心脏病(HHD)和肥厚型心肌病(HCM)进行自动分类。方法收集上海交通大学医学院附属仁济医院2017年1月至2021年12月收治的203名患者数据, 其中HHD患者53例, HCM患者121例, 正常对照(NC)组29例。所有患者均采用核磁共振仪采集native T1 mapping数据, 并采用多任务学习算法处理native T1 mapping数据, 以基于原始图像的Resnet 50模型为对照, 采用十折交叉、混淆矩阵和受试者特征(ROC)曲线验证各模型的分类性能。结果十折交叉验证结果显示, 与Resnet 50模型相比, MTL-1 024、MTL-64和MTL-all模型在曲线下面积(AUC)、准确率、敏感性和特异性等指标上均显示出更好的性能。在分类任务中, MTL-64模型的AUC(0.942 1)表现最佳, 而MTL-all模型的准确率达到了最高值(0.852 2)。在分割任务中, MTL-64模型的Dice系数(0.879 7)取得了最佳效果。混淆矩阵图表明MTL模型在整体性能上超越了基于原始图像的Resnet 50模型。且所有MTL模型的ROC曲线图明显高于原始图像输入Resnet 50模型。结论基于多任务学习的native T1 mapping图像对HHD和HCM的自动分类是有效果的。

Objective To automatically classify hypertensive heart disease(HHD)and hypertrophic cardiomyopathy(HCM)based on mul-titask learning algorithm using native T1 mapping images.Methods A total of 203 patients admitted to Ren Ji Hospital,School of Medicine,Shanghai Jiao Tong University from January 2017 to December 2021 were enrolled,including 53 patients with HHD,121 patients with HCM,and 29 patients with normal control(NC).Native T1 mapping images of all enrolled patients were acquired using MRI and processed by a multi-task learning algorithm.The classification performance of each model was validated using ten-fold crossover,confusion matrix,and receiver operator characteristic(ROC)curves.The Resnet 50 model based on the original images was established as a control.Results The ten-fold crossover validation results showed that the MTL-1024,MTL-64,and MTL-all models showed better performance in terms of area under the curve(AUC),accuracy,sensitivity,and specificity compared to the Resnet 50 model.In the classification task,the MTL-64 model showed the best performance in terms of AUC(0.9421),while the MTL-all model reached the highest value in terms of accuracy(0.8522).In the segmentation task,the MTL-64 model achieved the best results with the Dice coefficient(0.8797).The confusion matrix plot showed that the MTL model outperforms the Resnet 50 model based on the original image in terms of overall performance.The ROC graphs of all MTL models were significantly higher than the original image input Resnet 50 model.Conclusions Multi-task learning-based native T1 mapping images are effective for automatic classification of HHD and HCM.