基于Transformer模型的心音小波谱图识别

Recognition of Heart Sound Scalograms Based on Transformer Model

先天性心脏病(先心病)是严重威胁儿童健康的常见疾病,尽早进行先心病筛查对于该疾病的及时诊断和手术治疗十分重要,但这在医疗资源相对匮乏的偏远地区却难以实现。针对上述问题,为实现儿童先心病的智能化早期筛查,提出了一种基于复Morlet小波谱图与Transformer架构深度神经网络分类器的异常心音识别方法,其中小波谱图可以兼顾非平稳心音信号特征描述的时间分辨率与频率分辨率,而心音分类模型则是在经典ResNet50骨干网络的基础上嵌入了Transformer架构自注意力模块,可有效提升对时序信号谱图的特征提取能力。此外,还实现了基于Hilbert变换的心音包络检查,以及基于XMLRPC协议与Redis队列的并发心音识别服务部署,便于和前端机器人整合使用。实验测试显示识别准确率在现场心音数据集上达到87.5%,在PhysioNet 2016心音数据集上达到94.5%,表明该方法不仅在公开数据集上优于已有方法,即使是针对现场复杂环境下的心音识别任务也能取得较为理想的效果。

The congenital heart disease(CHD)is a common disease threatening children’s health severely.Early CHD screening is crucial for timely diagnosis and surgical treatment,but it is difficult to apply early screening in remote districts with relatively scarce medical resources.Aiming at the problem,a recognition method of abnormal heart sounds based on complex Morlet scalograms and a transformer-based deep neural network classifier is introduced to implement intelligent early CHD screening of children.The scalogram can balance the time resolution and the frequency resolution for the feature description of non-stationary heart sound signals.The classification model inserts a transformer-featured self-attention block into the classic ResNet50 backbone and improves the feature extraction ability for sequential signal spectrograms effectively.Besides,the envelope detection of heart sounds based on the Hilbert transform and the service deployment of a concurrent service of heart sound recognition based on XMLRPC and Redis are realized facilitating the integration of front-end robots.The experimental results show that the recognition accuracy for the working environment dataset reaches 87.5%and the value for the PhysioNet 2016 dataset is 94.5%,which indicates the proposed method here not only is superior to existing methods on the public dataset,but also can work well in complex realistic conditions.