Researchers seek help from machine learning methods to alleviate the enormous burden of reading radiological images for clinicians under the COVID- 19 pandemic. However, clinicians often feel reluctant to trust AI-based models because of its black-box characteristic and lack of proper explainability. This paper proposes an explainable attention transfer classification model based on the knowledge distillation network structure to automatically differentiate COVID-19, community acquired pneumonia (CAP) from healthy lungs with radiographic imaging. The proposed network structure can be divided into teacher network and student network based on the attention transfer direction. Firstly, the teacher network extracts global features and concentrates on the infection regions to generate attention maps. We propose a deformable attention module (DAM) to strengthen infection regions response and suppress noise in irrelevant regions with expanded reception field. Moreover, combining essential information in original input, attention knowledge transfers from teacher network to student network via an image fusion module. Trained with teacher network jointly, the student branch with weighted dense connectivity can focus on irregularly shaped lesion regions to learn discriminative features and improve network performance. Comprehensive experiments have been conducted on public chest X-ray and CT imaging datasets. The proposed architecture achieves state-of-art performance and improves the AI-based model's explainable ability by attention map, severity assessment, and prediction confidence.

中文翻译：

---

带有放射成像的COVID-19自动诊断：可解释的注意力转移到深层神经网络。

研究人员从机器学习方法中寻求帮助，以减轻在COVID-19大流行下临床医生读取放射图像的巨大负担。但是，由于其具有黑盒特性和缺乏适当的可解释性，临床医生通常不愿意信任基于AI的模型。本文提出了一种基于知识蒸馏网络结构的可解释的注意力转移分类模型，可以通过放射成像自动区分健康肺中的COVID-19，社区获得性肺炎（CAP）。根据注意力转移的方向，提出的网络结构可以分为教师网络和学生网络。首先，教师网络提取全局特征并专注于感染区域以生成注意力图。我们提出了一种可变形注意力模块（DAM），以增强感染区域的响应并在不相关区域中通过扩大接收场来抑制噪声。此外，将基本信息与原始输入相结合，注意力知识会通过图像融合模块从教师网络转移到学生网络。经过教师网络联合培训，具有紧密连接权重的学生分支可以专注于形状不规则的病变区域，以学习区分特征并提高网络性能。已经对公共胸部X射线和CT成像数据集进行了全面的实验。通过注意图，严重性评估和预测置信度，所提出的体系结构实现了最先进的性能并提高了基于AI的模型的可解释能力。