PURPOOSE: Convolutional neural networks (CNN) are deep learning algorithms which show tremendous ability for image classification. One caveat of this technology is the requirement for large datasets. Annotating and labeling of medical images is a difficult task which requires resources.
In this study, we examine the efficacy of creating synthetic images using generative adversarial network (GAN) for the classification of CT liver lesions using a CNN.
METHODS: This research was supported by the ISRAEL SCIENCE FOUNDATION (grant No. 1918/16).
Data augmentation is a common strategy used in computer vision to artificially increase the number of image samples. Classic data augmentation includes methods such as image rotation and flipping.
Generative adversarial networks (GANs) are a class of unsupervised deep learning algorithms, implemented by a system of two neural networks contesting with each other. One network generates candidates (generative) and the other evaluates them (discriminative). The generative network`s training objective is to "fool" the discriminator network by producing novel synthesized instances that appear to have come from the true data distribution. In the current study we used the Deep Convolutional GAN (DCGAN) architecture to synthesize images of CT liver lesions to be used for data augmentation. Lesions consisted of liver cysts, hemangiomas and metastases.
A nine layers CNN (three pairs of convolutional and max pooling layers, two fully connected layers and a Softmax activation layer) was used to classify the liver lesions.
The CNN’s sensitivity and specificity were compared between classic data augmentation and classic + synthetic data augmentation.
RESULTS: The non-augmented dataset included CT images of 182 liver lesions (53 cysts, 64 metastases and 65 hemangiomas). The classification performance using only classic data augmentation yielded 78.6% sensitivity and 88.4% specificity. By using classic and synthetic data augmentation the results increased to 85.7% sensitivity and 92.4% specificity.
CONCLUSION: The presented method for generating synthetic liver lesions images for data augmentation demonstrated improvement over classic data augmentation. Other radiological computer vision tasks may benefit from using synthetic data augmentation.
