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

https://doi.org/10.1016/j.compmedimag.2019.101662Get rights and content

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.

Introduction

Focal cortical dysplasia (FCD) is the malformation of the cortical development, which may be caused by reasons of cortical architecture or cytological abnormalities Kabat and Król (2012). It is the foremost cause of epilepsy in children and the third most significant cause in adults Lerner et al. (2009). Magnetic resonance imaging (MRI) is widely used in identifying FCD as it can provide images of soft tissue with high contrast and resolution. In Fig. 1, three T1-weighted MR images (i.e., one healthy and two FCD images) are shown. The regions enclosed by red rectangular boxes in Figs. 1 (b) and (c) show two typical MRI features of FCD: blurred gray matter (GM) - white matter (WM) boundary and increased cortical thickness. Manual identification of FCD lesion on MR images is a time-consuming and subjective task even for an experienced specialist due to the subtlety of FCD lesions and complexity of brain anatomic structure Rajan et al. (2009). In MR images, multiple tissue types (e.g., GM, WM, cerebrospinal fluid (CSF)) contribute to voxel/pixel intensities, which is known as partial volume effect, and it may also compromise the diagnosis of FCD. Therefore, it is important to develop a computerized system that can process brain MR images and detect subtle lesions automatically and objectively, to assist radiologists in analyzing images and making diagnosis.

Several techniques have been proposed for FCD detection in the literature. These techniques can be broadly divided into two categories: voxel-based morphometry (VBM), and surface-based classification (SBC). In the following, a brief review of techniques in these two categories is presented. Unless mentioned otherwise, T1-weighted MR images are assumed.

The VBM techniquesMechelli et al. (2005) typically normalize images to a standard stereotactic space, segment the normalized images into regions of interest (ROIs), smooth these ROIs, and finally performs a voxel-wise statistical analysis to highlight significant anatomical differences between patients and healthy groups. The techniques usually generate statistical feature maps that are used to detect FCD regions by choosing an optimum threshold. Colliot et al. Colliot et al. (2006) proposed to use GM concentration (GMC) map for FCD detection. The GMC map is calculated by segmenting a normalized brain into GM, WM, and CSF, smoothing the GM mask, and calculating z-score Kreyszig (1979) at each voxel in GM region. Note that z-score measures how many standard deviations a GMC value is above or below the mean of controls. A voxel is classified as FCD if its z-score value is greater than a predefined threshold. Similarly, Pail et al. Pail et al. (2012) used GMC map to detect FCD within the temporal pole in patients with mesial temporal lobe epilepsy. Wagner et al. Wagner et al. (2011) proposed to use GM-WM junction (GWJ) and GM extension (GME) maps for FCD detection. The calculation of GWJ includes brain normalization, segmentation, binarization to obtain GM-WM junctions, smoothing and comparison with normal database. The GME map is calculated using the GM segment of brain. Wong-Kisiel et al. Wong-Kisiel et al. (2018) further illustrated the effectiveness of GWJ and GME maps in FCD detection on a larger database of patients and controls. House et al. House et al. (2013) also used GWJ map, but based on T2-weighted MR images, to detect FCD regions.

The SBC techniques typically perform cortical reconstruction to obtain inner- and outer-cortical surfaces (using publicly available softwares such as FreeSurfer Laboratory for Computational (Neuroimaging)), extract features at each vertex, and finally classify lesion vertices using the machine learning methods, such as artificial neural network (ANN) and support vector machine(SVM). Different surface-based features and classification models have been proposed in these SBC techniques. Besson et al. Besson et al. (2008) calculated five features (i.e., cortical thickness, GM-WM blur, T1 hyper-intensity, sulcal depth, and curvature) at each vertex after cortical surface reconstruction, and performed both vertex-wise classification (using a bagged ANN) and cluster-wise classification (based on the statistic values of clusters). Hong et al. Hong et al. (2014) used similar surface features, but vertex-wise classification was performed using Fisher linear discriminant analysis and cluster-wise classification was based on the statistical moments of clusters. Ahmed et al. Ahmed et al. (2015) proposed to use bagged logistic regression to classify FCD vertices for “MRI-negative” patients. Adler et al. Adler et al. (2017) proposed additional “doughnut” features (calculated on a circle region of vertices on the inflated surface) in the vertex classification using ANN. Clustering was applied in the end to remove false positive FCD clusters. Tan et al. Tan et al. (2018) used surface features of both MRI and PET images, followed a two-step classification: a voxel-based SVM to maximize the sensitivity, and a patch-based classifier to remove the false positives. Jin et al. Jin et al. (2018) proposed a FCD detection technique similar to Adler et al. (2017), but with a larger database of images obtained from different epilepsy centers.

As observed, most techniques are based on either VBM or SBC. Typically, the VBM techniques use a few neurological features (e.g., GMC, GWJ, and GME), which are used for clinical diagnosis. Although, the VBM techniques have shown a good performance, these techniques are sensitive to artifacts including misalignment, misclassification, and differences in anatomical structures Ashburner (2009). On the other hand, the SBC techniques generate cortical surfaces, and calculate features that are robust to alignment and segmentation artifacts. These techniques are relatively efficient in FCD detection as they take into consideration the anatomical relationships across cortex Hong et al. (2014). But these techniques also have a high computational complexity because of the 3D surface reconstruction Riccelli et al. (2017). A further improvement may be possible by combining both voxel and surface features.

In recent years, deep learning techniques and especially convolutional neural networks (CNN) have shown great potentials in image classification and segmentation problems since they could learn optimal features automatically LeCun et al. (2015). Anthimopoulos et al. Anthimopoulos et al. (2016) proposed a lung pattern classification technique for interstitial lung disease based on a deep CNN architecture of 5 convolutional layers with 2×2 kernels, leaky rectifier linear unit (LReLU), and average pooling. The CNN architecture is trained and evaluated on non-overlapping image patches, and it shows superior classification performance over feature-based techniques (e.g., intensity, texture, LBP features). Pereira et al. Pereira et al. (2016) proposed a brain tumor segmentation technique based on a CNN architecture with 4-6 convolutional layers, 3×3 kernels, and LReLU. To the best of the authors knowledge, deep learning has not been applied to the detection of FCD. In this paper, we propose an automated FCD detection technique using a deep CNN architecture. The rest of this paper is organized as follows. In Section 2, the materials used in this work are introduced. The proposed methods are presented in Section 3, and the experimental results are reported in Section 4. Section 5 concludes this paper.

Section snippets

Materials

In this study, the patient group includes a retrospective cohort of 10 patients who underwent T1-weighted imaging on the 1.5T Siemens MRI scanner at the University of Alberta Hospital, with confirmed FCD lesions. The study was approved by the Alberta health services and University of Alberta research ethics board. The MR images were acquired using the T1-weighted magnetization-prepared rapid-acquisition gradient echo (MPRAGE) sequence (TR = 2130ms, TE = 3.91ms, and flip angle = 15) with an

Methods

The overall schematic of the proposed technique is shown in Fig. 2. It includes four modules: 1) preprocessing, 2) patch extraction, 3) deep learning classification, and 4) post-processing. Details of the proposed methods are presented in the following sections.

Experimental results

In this section, we first present the implementation of the proposed technique. The evaluation results of the proposed CNN model are then provided. Finally, the system-level evaluation is presented.

Conclusion

In this paper, an automated technique is proposed for FCD detection in T1-weighted MR images using a deep convolutional neural network. The proposed technique first performs preprocessing to align the input image with a standard brain atlas. The cortical patches are then extracted on axial slices. Each patch is paired and fed to a CNN classifier with 5 convolutional layers, 1 max pooling layer, and 2 fully-connected layers. Finally, the post-processing stage removes noises and correct missed

Acknowledgments

This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) (grant number: RGPIN-2014-05215). We also acknowledge the support of China Scholarship Council (CSC).

References (39)

  • V.S. Fonov et al.

    Unbiased nonlinear average age-appropriate brain templates from birth to adulthood

    NeuroImage

    (2009)
  • J.N. Mandrekar

    Receiver operating characteristic curve in diagnostic test assessment

    Journal of Thoracic Oncology

    (2010)
  • J. Kabat et al.

    Focal cortical dysplasia-review

    Polish journal of radiology

    (2012)
  • J.T. Lerner et al.

    Assessment and surgical outcomes for mild type I and severe type II cortical dysplasia: a critical review and the UCLA experience

    Epilepsia

    (2009)
  • D.H. Salat et al.

    Thinning of the cerebral cortex in aging

    Cerebral cortex

    (2004)
  • A. Mechelli et al.

    Voxel-based morphometry of the human brain: methods and applications

    Current medical imaging reviews

    (2005)
  • E.O. Kreyszig

    Advanced Engineering Mathematics

    (1979)
  • M. Pail et al.

    The role of voxel-based morphometry in the detection of cortical dysplasia within the temporal pole in patients with intractable mesial temporal lobe epilepsy

    Epilepsia

    (2012)
  • J. Wagner et al.

    Morphometric MRI analysis improves detection of focal cortical dysplasia type II

    Brain

    (2011)
  • Cited by (0)

    View full text