Classification of chest radiographs using depthwise separable convolution


Pneumonia x-ray images were classified using a neural network with depthwise separable convolutions. Each image was divided into three vertical regions and convolutions were applied to each region independently. The training model has less training parameters than a standard convolutional neural network (CNN) so that the tendency for overfitting and overall computation time is reduced. The trained network features a relatively high precision (ratio of true and predicted positives) and a significantly shorter training time than a conventional CNN. Our preliminary results indicate that the precision of the neural network was not greatly affected by the introduction of depthwise separable convolutions.