【摘要】 先天性心脏病(CHD)是婴幼儿死亡主要原因。基于人工智能的先天性心脏病诊断网络(CHDNet)是一种基于超声心动图视频的二分类模型，用于判别超声心动图视频是否包含心脏缺陷。现有的CHDNet模型表现出与医学专家相当甚至更好的判别性能，但它们在训练集之外的样本上的不可靠性已成为模型部署的关键瓶颈。而这是当前大多数基于AI诊断方法的共性问题。为了克服这一挑战，本文我们提出了两种基本机制——贝叶斯推理和动态神经反馈——分别用于衡量和提高人工智能诊断的可靠性。贝叶斯推理允许神经网络模型输出CHD判别的可靠性而不仅仅是单一的判别结果，而动态神经反馈是一个计算神经反馈单元，允许神经网络将知识从输出层反馈给浅层，使神经网络有选择地激活相关神经元。为了评估这两种机制的有效性，我们在包含三种常见CHD缺陷的4151个超声心动图视频上训练了CHDNet，并在1037个超声心动图视频的内部测试集和从其他心血管成像设备新收集的692个外部视频集上对其进行了测试。每个超声心动图视频对应于一位患者和一次就诊。我们在多种代表性神经网络架构上展示了贝叶斯推理获得的可靠性如何解释和量化神经网络内部与外部测试集之间的性能显著差异，以及尽管输入被噪声破坏或使用外部测试集时，设计的反馈单元如何帮助神经网络保持高精度和可靠性。更多还原

【Abstract】 Congenital heart disease(CHD) is the leading cause of infant death. An artificial intelligence(AI)-based CHD diagnosis network(CHDNet) is an echocardiogram video-based binary classification model that judges whether echocardiogram videos contain heart defects. Existing CHDNets have shown performances comparable to or even better than medical experts, but their unreliability on cases outside of the training set has become the main bottleneck for their deployment. This is a common problem for most AI-based diagnostic approaches. Here, to overcome this challenge, we present two essential mechanisms-Bayesian inference and dynamic neural feedback-to respectively measure and improve the diagnostic reliability of AI. The former easily makes the neural network output its reliability instead of a single prediction result, while the latter is a computational neural feedback cell that allows the neural network to feed knowledge from the output layer back to the shallow layers and enables the neural network to selectively activate relevant neurons. To evaluate the effectiveness of these two mechanisms, we trained CHDNets on 4151 echocardiogram videos containing three common CHD defects and tested them on an internal test set of 1037 echocardiogram videos and an external set of 692 videos that were newly collected from other cardiovascular imaging devices. Each echocardiogram video corresponds to a unique patient and a unique visit. We demonstrate on various neural network architectures how the reliability obtained by Bayesian inference interprets and quantifies the significant performance difference between internal and external test sets of neural networks, and how the devised feedback cell helps the neural networks to maintain high accuracy and reliability, despite the input being corrupted by noise or when using an external test set.更多还原