Deep neural network with generative adversarial networks pre-training for brain tumor classification based on MR images. (March 2020)

• Record Type: Journal Article
• Authors: Ghassemi, Navid; Shoeibi, Afshin; Rouhani, Modjtaba
• Is Part Of: Biomedical signal processing and control. Volume 57(2020)
• Publication Date: 2020-03
• Subjects: Magnetic resonance imaging (MRI) -- Deep neural networks -- Generative adversarial network (GAN) -- Brain tumor classification; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2019.101678
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Presenting an unsupervised pretraining method to overcome overfitting using GAN. The unsupervised pretraining allow using of similar unlabeled datasets. This method can be used for training complex networks on small datasets. State-of-the-art results are obtained on a brain tumor classification dataset. Abstract: In this paper, a new deep learning method for tumor classification in MR images is presented. A deep neural network is first pre-trained as a discriminator in a generative adversarial network (GAN) on different datasets of MR images to extract robust features and to learn the structure of MR images in its convolutional layers. Then the fully connected layers are replaced and the whole deep network is trained as a classifier to distinguish three tumor classes. The deep neural network classifier has six layers and about 1.7 million weight parameters. Pre-training as a discriminator of a GAN together with other techniques such as data augmentations (image rotation and mirroring) and dropout prevent the network from overtraining on a relatively small dataset. This method is applied to an MRI data set consists of 3064 T1-CE MR images from 233 patients, 13 images from each patient on average, with three different brain tumor types: meningioma (708 images), glioma (1426 images), and pituitary tumor (930 images). 5-Fold cross-validation is used to evaluate the performance of overall design, achieving the highest accuracy as compared to state-of-art methods.