Custom convolutional neural network with data augmentation to predict Pneumonia COVID19

Satoto, B. D.; Utoyo, M. I.; Rulaningtyas, R.; Koendhori, E. B.

Coronavirus is named because of the structure of the crown on its body surface. The effect of the Coronavirus for sufferers is a disturbance in the respiratory system. Several days later, the disruption from the lung infection got worse. To identify the cause of this disease, the doctor performs a computed tomography scan and manually observes the changes that occur in the lungs through an X-ray. Image identification using machine learning is the latest trend these days to assist medical analysts. If the number of patients treated is large enough, this is very helpful in the analysis. The choice of Convolutional Neural Network is due to the many architectural algorithms being developed at this time. This method works with multiple layers. But the drawback is that the computation time for the training process takes a long time. The purposed way in this research is a custom layer using 18-34 layers. There is four class in the test, namely Normal lung conditions, COVID19, bacterial pneumonia, and viral pneumonia. Data augmentation is used to add variation to data. The results showed that the method offered could be used to identify pneumonia with an average identification accuracy of 98.7% - 100%. The average value of error the MSE 18-34 layer is 0.0539, RMSE 0.1981, and MAE 0.0319. The average consumption time for the training process is 2.25 seconds. The best accuracy calculation is obtained at 34 layers with the Adaptive Moment Estimation optimizer with a computation time of around 1 minute 48 seconds. © 2020 IEEE.