Automated detection of focal cortical dysplasia using a deep convolutional neural network

Abstract

Focal cortical dysplasia (FCD) is one of the commonest epileptogenic lesions, and is related to malformations of the cortical development. The findings on magnetic resonance (MR) images are important for the diagnosis and surgical planning of FCD. In this paper, an automated detection technique for FCD is proposed using MR images and deep learning. The input MR image is first preprocessed to correct the bias field, normalize intensities, align with a standard atlas, and strip the non-brain tissues. All cortical patches are then extracted on each axial slice, and these patches are classified into FCD and non-FCD using a deep convolutional neural network (CNN) with five convolutional layers, a max pooling layer, and two fully-connected layers. Finally, the false and missed classifications are corrected in the post-processing stage. The technique is evaluated using images of 10 patients with FCD and 20 controls. The proposed CNN shows a superior performance in classifying cortical image patches compared with multiple CNN architectures. For the system-level evaluation, nine of the ten FCD images are successfully detected, and 85% of the non-FCD images are correctly identified. Overall, this CNN based technique could learn optimal cortical (texture and symmetric) features automatically, and improve the FCD detection.