Effect of augmented datasets on deep convolutional neural networks applied to chest radiographs

Abstract

Aim: To evaluate the effect of augmented training datasets in a deep convolutional neural network (DCNN) used for detecting abnormal chest radiographs.

Materials and methods: Chest radiographs were corrected to conform to a DCNN dataset, with 288 abnormal and 447 normal radiographs. The radiographic images were divided into training and validation sets (441, 60%), and a test set (294, 40%). The training and validation sets were augmented to generate a total of 12,789 training and validation images. The augmentation consisted of operations such as rotation, horizontal and vertical flipping, Gaussian blur, and brightness variation, either alone or combined. The DCNN performed binary classification of the images as being abnormal or normal chest radiographs, and accuracy was used as measure to assess the model performance.

Results: The accuracy of the DCNN trained with the augmented dataset tended to be higher than that of the DCNN trained with the non-augmented dataset. The augmented datasets combining rotation and horizontal flipping had a high accuracy of 0.91, showing the highest accuracy among the applied augmentation techniques and combinations.

Conclusion: Augmentation of training datasets can improve the performance of DCNN for radiographic image classification depending on the applied augmentation technique.