人工智能辅助床旁超声诊断肾创伤的基础研究

目的探讨基于卷积神经网络（CNN）构建的人工智能辅助诊断模型对肾钝性创伤超声诊断的应用价值。 方法建立不同程度动物肾创伤模型，通过床旁超声仪采集正常肾及创伤肾超声图片，分成训练集及测试集，根据造模位置和超声造影结果，手动勾画出肾轮廓，采用3折交叉验证进行分类训练及测试。绘制受试者工作特征（ROC）曲线，计算人工智能辅助诊断模型的敏感度、特异度、准确性和曲线下面积（AUC）。 结果采集正常肾图片共1737张，各级别创伤肾图片共2125张，经过对测试集的验证，该模型可自动对肾创伤有无进行分类，对肾创伤诊断的平均敏感度为73%、平均特异度为85%、平均准确性为79%、AUC为0.80，诊断价值较高。 结论基于CNN构建的深度学习模型辅助床旁超声仪在诊断肾创伤有无分类中取得了较满意的结果。

Artificial intelligence aided point of care ultrasound for diagnosis of renal injury: a pilot study

ObjectiveTo explore the application value of artificial intelligence aided diagnosis model based on convolutional neural network (CNN) in ultrasonic diagnosis of blunt renal trauma. MethodsRabbits were used to simulate different grades of renal trauma model of renal trauma of different degrees was established. The ultrasonic images of the normal kidney and renal trauma were collected by point of care ultrasound (POCUS) and divided into either a training or a test cohort. According to the modeling position and contrast-enhanced ultrasound results, the renal contour was manually drawn and classified for training, followed by 3-fold cross-validation testing. The sensitivity, specificity, accuracy, and area under curve (AUC) of the artificial intelligence aided diagnosis model were calculated. ResultsA total of 1737 images of the normal kidney and 2125 images of traumatic kidney were collected. After the verification of the test set, the model can automatically classify the presence or absence of renal trauma. The average sensitivity for renal trauma diagnosis was 73%, the average specificity was 85%, the average accuracy was 79%, and the AUC was 0.80. ConclusionThe deep learning assisted POCUS model constructed based on CNN has achieved satisfactory results in the diagnosis and classification of renal trauma.