Differentiation of Cytopathic Effects (CPE) induced by influenza virus infection using deep Convolutional Neural Networks (CNN)

Author summary Observation of cytopathic effects (CPE) induced by virus infection is a practical method to determine the prsence of viruses in the clinical specimens. However, CPE observation is labor-intensive and time-consuming because it requires medical examiner to inspect cell morphology changes for a period of time. Here, Convolutional Neural Networks (CNN) was applied to improve the disadvantage of CPE observation by using influenza virus as an example. To reduce the requirement for large image input of every clinical test, small amount of data was used to train our CNNs model without transfer learning and the trained model was examined with testing image data taken at 25hr post virus infection. The recognition of testing data shows that the model can identify CPE at 25hr and the high specificity of the model can differentiate the CPE induced by influenza viruses from those by other non-influenza viruses. The limit of our model was further examined by more experimental data of influenza-induced and mock-infected images, and the result shows our model can detect the slight changes at the initial stage of CPE development. Hence, our deep CNN model can significantly shorten the timing required to identify virus-induced cytopathic effects. Cell culture remains as the golden standard for primary isolation of viruses in clinical specimens. In the current practice, researchers have to recognize the cytopathic effects (CPE) induced by virus infection and subsequently use virus-specific monoclonal antibody to confirm the presence of virus. Considering the broad applications of neural network in various fields, we aimed to utilize convolutional neural networks (CNN) to shorten the timing required for CPE identification and to improve the assay sensitivity. Based on the characteristics of influenza-induced CPE, a CNN model with larger sizes of filters and max-pooling kernels was constructed in the absence of transfer learning. A total of 601 images from mock-infected and influenza-infected MDCK cells were used to train the model. The performance of the model was tested by using extra 400 images and the percentage of correct recognition was 99.75%. To further examine the limit of our model in evaluating the changes of CPE overtime, additional 1190 images from a new experiment were used and the recognition rates at 16 hour (hr), 28 hr, and 40 hr post virus infection were 71.80%, 98.25%, and 87.46%, respectively. The specificity of our model, examined by images o