Multiple sclerosis identification by convolutional neural network with dropout and parametric ReLU

• Published in: Journal of computational science Vol. 28; pp. 1 - 10
• Main Authors: Zhang, Yu-Dong, Pan, Chichun, Sun, Junding, Tang, Chaosheng
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.09.2018
• Subjects: Convolutional neural network; Data augmentation; Dropout; Leaky rectified linear unit; Multiple sclerosis; Parametric rectified linear unit; Retention probability; Multiple sclerosis; Convolutional neural network; Leaky rectified linear unit; Data augmentation; Parametric rectified linear unit; Retention probability; Dropout
• ISSN: 1877-7503; 1877-7511
• DOI: 10.1016/j.jocs.2018.07.003

•Our improved convolutional neural network combined the parametric rectified linear unit and dropout techniques.•A 10-layer deep convolutional neural network was established, with 7 convolution layer and 3 fully connected layers.•Our method achieved a sensitivity of 98.22%, a specificity of 98.24%, and an accuracy of 98.23%.•Dropout increases the accuracy by 0.88% compared to not using dropout.•PReLU helped increase the accuracy by 1.91% compared to using ordinary ReLU. Multiple sclerosis is a condition affecting brain and/or spinal cord. Based on deep learning, this study aims to develop an improved convolutional neural network system. We collected 676 multiple sclerosis brain slices and 681 healthy control brain slices. Data augmentation was used to increase the size of training set. Our improved convolutional neural network combined the parametric rectified linear unit (PReLU) and dropout techniques. Finally, a 10-layer deep convolutional neural network was established, with 7 convolution layer and 3 fully connected layers. The retention probabilities of three dropout layers are set as 0.4, 0.5, and 0.5, respectively. Our method achieved a sensitivity of 98.22%, a specificity of 98.24%, and an accuracy of 98.23%. The dropout helped increase the accuracy by 0.88% compared to not using dropout. PReLU helped increase the accuracy by 1.92% compared to using ordinary ReLU, and by 1.48% compared to using leaky ReLU. This proposed method is superior to four state-of-the-art approaches.