基于Transformer的新生儿肺部超声图像诊断

The Diagnosis of Neonatal Lung Disease Based on Ultrasound Images Using Transformer

摘要: 新生儿肺超声是近几年发展起来的新技术，也是对新生儿肺部疾病进行观察与诊断的关键技术。本研究研讨基于深度学习的分类模型，对新生儿肺部超声图像自动进行阴性与阳性的分类，回顾性搜集了2020年在上海市第一妇婴保健院出生的新生儿的肺部超声图像，基于Pytorch的Vision Transformer框架，并结合卷积神经网络构建模型。在多次实验后，测试集的准确率达到了95%，敏感性为98%，特异性、精确率均达到了93%，ROC曲线的AUC达到了99%。通过卷积与Transformer相结合构建的分类模型，对新生儿肺部超声图像的分类效果较好，有助于医生对患者的进一步诊断与治疗。

Abstract: Neonatal lung ultrasound is a new technique that has been developed in recent years, and it is also a key technology to observe and diagnose neonatal pulmonary diseases. This study examines a classification model based on deep learning to automatically classify negative and positive lung ultrasound images of newborns. We collected the ultrasonic images of the lungs of newborns born in the Shanghai First Maternity and Infant Hospital in 2020. Based on the Pytorchand Vision Transformer, and combined with convolution neural network, a model is built. After many experiments, the accuracy of the test set is 95%, the sensitivity is 98%, the specificity and accuracy are 93%, and the AUC of the ROC is 99%. The classification model constructed by the combination of convolution and Transformer has a good classification effect on neonatal lung ultrasound images, which is helpful for doctors to further diagnose and treat patients.