Superior sensitivity to focal cortical dysplasia obtained by a multivariate analysis of MRI and PET image features
Hosung Kim1, Yee-Leng Sung Tan2, Tarik Tihan3, Anthony James Barkovich1, Duan Xu1, and Robert C Knowlton2

1Radiology & Biomedical Imaging, University california San francisco, San Francisco, CA, United States, 2Neurology, University california San francisco, SAN FRANCISCO, CA, United States, 3Pathology, University california San francisco, SAN FRANCISCO, CA, United States


Focal cortical dysplasia (FCD) is an epileptogenic developmental malformation. Identification of this lesion can lead to a successful surgery. We propose to analyze a combined feature-set extracted from MRI and PET. Studying 29 FCD patients and 23 controls, classification using the combined MRI and PET features demonstrated superior performance to the analysis of MRI as it resulted in a lower false positive (FP) rate in controls (1.3% lower) and a higher sensitivity in FCD (7% higher). Analysis of the combined MRI and PET revealed a larger FP rate in FCD compared to MRI-only, suggesting the presence of extralesional pathology.

Background and Purpose

Focal cortical dysplasia (FCD) is an epileptogenic developmental malformation, characterized by a range of abnormalities, including cortical dyslamination, cytomegalic dysmorphic neurons, and balloon cells. FCD is a common cause of intractable epilepsy in children, and the most common histopathology in extratemporal lobe epilepsy 1. Identification of this lesion can lead to a successful epilepsy surgery 1. Because the scalp electrochephalogram has a low spatial resolution, anatomical MRIs such as t1- and t2-weighted images have been widely used to improve detection of epileptogenic lesions to aid localization of focal epilepsy. Combination of a pattern classification with quantitative imaging features representing FCD lesions including cortical thickening, blurring of gray matter (GM) – white matter (WM) junction, and hyperintense GM signal, has dramatically enhanced the performance of the FCD identification 2. A further improvement of the sensitivity, however, is required for clinical practice 2. Recent evidence has suggested positron emission tomography (PET) may complement the sensitivity of MRI, in particular for cases with small or more subtle lesions, including those that are cryptogenic, i.e. “MRI-negative” 3. We propose a framework to analyze a set of features extracted from both MRI and PET to address the need for improved diagnostic sensitivity. We used patients with temporal lobe epilepsy as controls because it is difficult to recruit healthy controls in most clinical environments.


We studied 29 FCD patients and 23 TLE controls (Table 1). No FCD patients presented neurological abnormalities other than FCD or dual pathology (e.g., mesiotemporal sclerosis). Individuals underwent T1-weighted MRI scanning and FDG-PET scanning. The MR images were first processed to correct intensity non-uniformity and to normalize the intensity range. They were then registered to the MNI-ICBM template. Cortical surfaces representing the WM-GM and GM-CSF interface were extracted from the registered MRI 4. The PET image and MR image were co-registered within an individual using a mutual information-based rigid-body transformation (Fig1A). To avoid data interpolation related to registration, features were computed after surfaces were transformed back into each image’s native space. The following features were computed using cortical surfaces: cortical thickness, sulcal depth (small lesions are known to be located in the bottom of a deep sulcus), relative intensity (hyperintensity in GM), gradient along surface-normal direction (i.e., blurring), and PET intensity (normalized using total WM intensity and CSF intensity in the lateral ventricle). Each feature was normalized: (1) across all the surface points within a given individual (intra-subject normalization); and (2) feature values at a given point were normalized with respect to the distribution of TLE controls (inter-subject normalization). Classification was performed using Fisher linear discriminant analysis that found the optimal weights for a linear combination of features to achieve maximal separation between classes. We trained and cross-validated the classifier using a leave-one-out strategy, by which a patient is classified based on data of all patients other than that patient.


An example of classification is shown in Fig1B. For both analyses of MRI only and combined MRI and PET features, we found the best trade-off between the sensitivity and specificity at the posterior probability of 91% in the classifier as we obtained 100% detection rate of true positive (TP) FCD lesions (29/29) and the lowest false positive (FP) rates (Table 2). Classification using the combined MRI and PET features demonstrated superior performance to the analysis of MRI as it resulted in a lower FP rate in TLE controls (1.3% lower). On the other hand, use of the combined MRI and PET revealed a larger FP rate in FCD patients compared to using MRI only. This pattern was kept even at the highest posterior probability (99%) where very low FP rate was found in TLE controls (0.6%), suggesting the presence of extralesional pathology. At the maximum posterior probability, the combined analysis of MRI and PET also showed higher sensitivity to FCD lesions than the MRI only (86% vs. 79%).


Our analysis of the clinical utility of imaging features for the detection of FCD lesions showed that superior sensitivity and specificity of the classification were obtained by combining PET with MRI, when compared to the conventional analysis of MRI only. Our analysis further demonstrated the presence of extra-lesional brain abnormalities in patients with FCD, which is supported by a previous report 2. Use of TLE controls in our study yielded classification performance parallel to using healthy controls as reported in a previous analysis 2, suggesting the possibility to exploit more widely available databases. An additional classification step may be required to remove the relatively large amount of FP clusters in FCD patients 3,5.


No acknowledgement found.


1. Blümcke I et al., Epilepsia 2011; 2. Hong SJ et al., Neurology 2014; 3. Guerrini R et al., Epilepsia 2015; 4. Kim JS et al., Neuroimage 2005; 5. Besson P et al., MICCAI 2008.


Fig 1. A: PET-MRI coregistration. The FCD lesion in red. B: Imaging features mapped on surface & classification result (TP: blue; FP: white).

Table 1. Demographics

Table 2. Classification results

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)