用于新冠肺炎CAR的类残差CNN-LSTM

Residual-like CNN-LSTM for Computer-Aided Recognition of COVID-19

新型冠状病毒肺炎目前已成为全球性的重大公共卫生事件。反转录·聚合酶链反应检测是检测新型冠状病毒肺炎的黄金手段,但从经济角度与效率角度来说,采用基于图像识别技术的计算机辅助诊断则是另一种行之有效的辅助检测手段,提出了一种类残差CNN-LSTM神经网络,针对串行结构卷积神经网络,采用类似于残差网络的思想提取图像的多级抽象特征并使用长短期记忆网络对其进行融合后识别;针对并行卷积神经网络,使用长短期记忆网络融合来自不同结构卷积神经网络的特征后进行识别。上述方法在加州大学开源的数据集上进行了验证,取得了Recall为0.9655,F1-score为0.8819,accuracy为87.25%,AUC为90.72的识别结果,相较于传统结构的卷积神经网络,各项性能指标提高了2~10个百分点。

COVID-19 has now become a major global public health event. Reverse transcription and polymerase chain reaction detection are the golden methods for detecting novel coronavirus pneumonia, but from the economic and efficiency perspective, the use of computer-aided diagnosis based on image recognition technology is another effective auxiliary detection method. In this paper, a kind of residual CNN-LSTM neural network is proposed. Aiming at the serial structure convolutional neural network, the idea similar to the residual network was used to extract the multi-level abstract features of the image and the long and short-term memory network was used to recognize them after fusion. For parallel convolutional neural networks, long and short-term memory networks were used to fuse features from different convolutional neural networks for recognition. This method was verified on the open source data set of the University of California, and the Recall was 0.9655,the F1-score was 0.8819,the accuracy was 87.25%,and the AUC was 90.72. Compared with the traditional structure of the convolutional neural network, the recognition results Performance indicators have increased by 2 to 10 percentage points.