Automated morphometric magnetic resonance imaging analysis for the detection of periventricular nodular heterotopia


  • [Correction added after online publication 18-Jan-2013: Dr. Pascher's name has been updated.]

Address correspondence to Hans-Jürgen Huppertz, Swiss Epilepsy Centre, Bleulerstrasse 60, CH-8008 Zürich, Switzerland. E-mail:


Purpose:  To describe a novel magnetic resonance imaging (MRI) postprocessing technique for the detection of periventricular nodular heterotopia (PNH) and to evaluate its diagnostic value. The method is a further development of voxel-based morphometric analysis with focus on a region of interest around the lateral ventricles to increase the sensitivity and specificity for automated detection of abnormally located gray matter in this area.

Methods:  T1-weighted MRI volume data sets were normalized and segmented in statistical parametric mapping (SPM 5 software), and the distribution of gray matter was compared to a normal database. As a new approach, individual masks derived from segmentation of the lateral ventricles were used to restrict the search for ectopic gray matter to the periventricular area. PNH were automatically detected by localizing the maximum deviation from the normal database in this area, provided that the z-score exceeded a certain threshold. The optimal z-score threshold for maximum sensitivity and specificity was determined by a receiver operating characteristic (ROC) curve analysis. The method was applied in 40 patients with PNH and 400 controls.

Key Findings:  PNH were detected in 37 of 40 patients, and false positives were found in 34 of 400 controls, amounting to 92.5% sensitivity and 91.5% specificity. In 17 of the patients in whom PNH could be identified, these lesions had been overlooked in the past, and in 8 patients even in the high-resolution MRI subsequently used for postprocessing.

Significance:  The results suggest that automated morphometric MRI analysis with focus on ectopic gray matter in the periventricular areas facilitates the evaluation of MRI data and increases the sensitivity for the detection of PNH.

Periventricular nodular heterotopia (PNH) is regarded as a malformation of cortical development due to abnormal neuronal migration (Barkovich et al., 2005). Ectopic gray matter (GM) is located in nodules along the walls of the lateral ventricles close to the subependymal germinal matrix. PNH is often associated with epilepsy beginning in the second decade of life, sometimes also with mild mental retardation and neurologic deficits. Less commonly, PNH is part of complex brain malformation and found in patients with severe epilepsy and mental retardation (d’Orsi et al., 2004; Battaglia et al., 2006; Parrini et al., 2006; Srour et al., 2011).

In magnetic resonance imaging (MRI), PNH appears as round or ovoid nodules that are isointense with mature GM in all imaging sequences and protrude slightly into the ventricular lumen, resulting in an irregular ventricular outline (Barkovich & Kuziecky, 2000). In contrast to nodules in tuberous sclerosis, PNH exhibits no gadolinium enhancement or calcification. In principle, visual detection in good-quality MRI images should not pose difficulties, especially if the predilection sites are scrutinized (e.g., trigones, temporal and occipital horns of the lateral ventricles). In clinical reality, however, the detection depends very much on the attention and diligence of the investigator, and particularly for small and solitary lesions the recognition can be difficult.

Morphometric MRI analysis has proven to be helpful in the evaluation of MRI data by highlighting aberrant neuronal tissue extending beyond the normal cortical ribbon (i.e., abnormal deep sulci or completely ectopic GM). Following the principles of voxel-based morphometry (VBM) (Ashburner & Friston, 2000; Ashburner et al., 2003) and starting from a T1-weighted volume data set, this technique primarily includes a normalization and GM segmentation using SPM (statistical parametric mapping software, Wellcome Trust Centre for Neuroimaging, London, United Kingdom; and the detection of abnormally located GM by comparison with a normal database (Kassubek et al., 2002; Wilke et al., 2003; Huppertz et al., 2005). Herein we introduce a further development of this approach. By using masks based on the individual space of the cerebrospinal fluid in the investigated patient, the search for aberrant GM is focused and limited to a region of interest around the lateral ventricles in order to increase the sensitivity and specificity for automated detection of PNH. In our study, we present first results of this method in patients with PNH compared to healthy controls.


Patients and controls

At the Swiss Epilepsy Center in Zürich, a register has been maintained for all patients receiving an MRI since January 2006. In 55 (1.48%) of 3,719 patients included until December 2011, PNH was diagnosed according to typical MRI criteria (i.e., isointensity to GM in all imaging sequences, no gadolinium enhancement, no sign of calcification in T2* sequence, formation in nodules along the walls of the lateral ventricles). Fifteen patients had to be excluded from this study because no T1-weighted volume data set was available (12/15), or MRI postprocessing was not possible because of major cerebral abnormalities (microcephaly, gross ventricle enlargement) (3/15).

All remaining 40 patients (19 female, 21 male, mean age of 32.8 ± 14.4 years, range 4.3–65.8 years) had a high-resolution MRI including an unenhanced T1-weighted volume data set of 1 mm3 voxel size. The majority (25/40) of patients were investigated at a Philips Achieva 3T scanner (Philips, Amsterdam, The Netherlands) using a three-dimensional Turbo Field Echo sequence (repetition time 8.1 msec; echo time 3.7 msec; flip angle 8 degrees; field of view 256 × 256 mm, slab thickness 160 mm; sagittal acquisition), eight patients at a Siemens Sonata 1.5T scanner (Siemens, Erlangen, Germany), seven patients at a GE 3T scanner (General Electric Healthcare, Chalfont St. Giles, United Kingdom), and four patients at a Philips Intera 3T scanner.

The control group consisted of 400 healthy individuals (208 female, 192 male, mean age 32 ± 11.7 years, range 16.5–74.5 years), extracted by randomization from a pool of controls whose MRI data sets have been used in the past for creating scanner-specific normal databases for MRI postprocessing. The inclusion criteria were the following: no diagnosis of epilepsy or other neurologic diseases, no pathologic findings in the MRI after thorough inspection by a radiologist. The unenhanced T1-weighted volume data sets of these controls were also acquired at different scanners (i.e., Siemens Trio 3T [180/400], Siemens Sonata 1.5T [52/400], Siemens Symphony 1.5T [45/400], Siemens Skyra 3T [24/400], Siemens Vision 1.5T [16/400], Philips Achieva 3T [83/400]). Overall, 113/400 images were acquired at 1.5 Tesla and 287/400 at 3 Tesla.

MRI postprocessing

The MRI postprocessing was based on standard procedures available within SPM5 (e.g., normalization, segmentation) and on additional simple computations and filters (e.g., calculation of a difference image, conversion to a binary image, masking, smoothing). It was fully automated using a MATLAB (MathWorks, Natick, MA, U.S.A.) batch script and comprised the following steps (see also Fig. 1; the numbers within the figure correspond to these processing steps). The first steps, that is, for the creation of the so-called “extension image,” have already been described in the past (Kassubek et al., 2002; Wilke et al., 2003; Huppertz et al., 2005).

Figure 1.

Overview of the image processing steps: (1) normalization and bias correction of the T1 input image, (2) segmentation of the T1 image, (3) smoothing of the segmented GM image, (4) comparison with a normal database, (5) calculation of a z-score image, the so-called “extension image,” which highlights ectopic gray matter all over the brain, (6) segmentation of the lateral ventricles, (7) smoothing of the “lateral ventricle image” to define a region of interest around the lateral ventricles, (8) calculation of the final “PNH image” by multiplying the “extension image” and the “smoothed lateral ventricle image” in order to highlight only abnormally located gray matter in the vicinity of the lateral ventricles.

1 and 2: Normalization, intensity correction, and simultaneous segmentation

SPM5 includes a probabilistic framework (called “unified segmentation”) whereby image registration, tissue classification, and bias correction are integrated within the same generative model (Ashburner & Friston, 2005). With use of this framework, the three-dimensional T1 data set of each subject is normalized to the standard brain of the Montreal Neurological Institute (MNI) included in the SPM distribution, segmented into different brain compartments, that is, GM, white matter (WM), and cerebrospinal fluid (CSF), and is simultaneously corrected for small intensity inhomogeneities (using default SPM5 parameters).

3: Smoothing of the GM image

The GM image resulting from segmentation is smoothed by a Gaussian kernel of 6-mm full width at half maximum (FWHM). In the smoothed GM image, each voxel encodes the average concentration of GM from around the voxel at the corresponding position in the original structural MR image.

4: Comparison with a normal database

To compensate for the variability of GM distribution in the normal population, the smoothed GM image of the investigated subject is compared with a normal database. This normal database has been created in the past from three-dimensional T1 images of 150 controls measured at five different 1.5 and 3T MR scanners. This population is different from the controls for this study and has been described in detail elsewhere (Huppertz et al., 2008a,b). The T1 images of the normal database have been processed in the same way as described in steps 1–3, and their smoothed GM images have been averaged. The resulting mean GM image of the normal database is subtracted voxel by voxel from the smoothed GM image of the subject investigated.

5: Calculation of the “extension image”

The smoothed GM images of the normal database have also been used to calculate a “standard deviation (SD) image” providing standard deviations of the normal database for all voxels. The difference image resulting from step 4 is divided by this SD image of the normal database to obtain the so-called “extension image” with z-score normalized data. In order to avoid outlier values at the border of the standard brain where only few subjects contribute to the normal database and its SD, the SD image has previously been smoothed by using a fixed Gaussian kernel of 6-mm FWHM. In the extension image, those brain areas appear bright where GM extends abnormally into the white matter or is abnormally located as compared with the normal database.

6: Segmentation of the lateral ventricles

The normal database was also employed to create a binary mask for lateral ventricles. The CSF images resulting from the segmentation of the 150 T1 images were averaged to get a mean CSF image. In this image, the lateral ventricles were manually delineated in their outermost boundaries. The final “lateral ventricle mask” was formed from all voxels of the mean CSF image falling within these boundaries and exceeding a lower threshold of 0.1, that is, where the average CSF probability of the 150 controls was at least 10%. However, this “lateral ventricle mask” is only an intermediate step in the creation of an individual ventricle mask for the investigated subject. To delineate the extent of the lateral ventricles in an individual subject, his CSF image derived from the segmentation in step 2 is multiplied with this “lateral ventricle mask.”

7: Smoothing of the “lateral ventricle image”

The resulting “lateral ventricle image” of the individual subject is smoothed by a Gaussian kernel of 6-mm FWHM and then clipped by thresholding at 0.01 to create a binary mask, which comprises a space of about 1 cm around the lateral ventricles. This defines the region of interest for the subsequent search for PNH lesions.

8: Calculation of the PNH image

In the last step, the extension image and the smoothed lateral ventricle image are multiplied to get the final “PNH image,” which highlights abnormally located GM only in the vicinity of the lateral ventricles.

Using a predefined z-score threshold, the PNH images are searched for regions where the distribution of GM differs significantly from the normal database and which therefore could represent PNH lesions. Suspicious alterations are then highlighted both in the extension image and in the original T1 image. Figure 2 shows the results of the MRI postprocessing in Fig. 1, that is, the original T1-weighted image and the corresponding extension image. The crosshairs are centered on the voxel with the highest z-score found in the PNH image, and beneath the crosshairs the extension image displays a bright spot that corresponds to ectopic GM in the border of the right lateral ventricle as shown in the T1 image.

Figure 2.

Result of the example image processed in Fig. 1: automated detection of PNH in the right frontocentral region by localizing the voxel with the highest z-score in the PNH image, which exceeds a certain predefined z-score threshold. The crosshairs show the position in axial, coronal, and sagittal slices of the T1 image (upper row) and the corresponding “extension image” (lower row).

Evaluation of results

To determine the z-score threshold that provides the highest sensitivity and specificity for the detection of ectopic GM and differentiates best between true PNH and false-positive results, a receiver operating characteristic (ROC) curve analysis was employed. The ROC curve is a graphical plot of the true-positive rate versus false-positive rate, and shows the correlation between sensitivity and specificity across a series of cutoff points for a dichotomous test (Kumar & Indrayan, 2011). The area under the curve (AUC) is a combined measure of sensitivity and specificity for assessing the performance of the diagnostic test (possible range: 0.5–1; best performance at 1). The point on the ROC curve closest to the corner of the diagram that represents the theoretically best possible performance (i.e., both sensitivity and specificity of 100%) indicates the optimal threshold for differentiation of true positives and false positives.

For each of the 40 PNH patients and the 400 controls, the highest z-score in the PNH image was determined. The results were used as input for an ROC analysis as described above. By variation of the z-score threshold between its extrema, that is, the lowest and the highest values found in the study population, the optimal cutoff for differentiation of true positives (i.e., ectopic GM in PNH patients) and false positives (i.e., any findings in the healthy controls) was determined. The analysis was performed by using the MATLAB scripts “roc.m” (for the calculation of the ROC curve) and “partest.m” (for the calculation of the optimal trade-off) created by G. Cardillo (Mathworks File Exchange:


Within the group of patients with PNH included in this study, 23 had unilateral and 17 had bilateral PNH, and 19 had isolated and 21 multiple nodules. In eight patients, additional epileptogenic lesions such as complex cerebral malformation (4), focal cortical dysplasia (1), polymicrogyria (1), mesial temporal lobe sclerosis (1), and hypothalamic hamartoma (1) were found. On average, each patient had undergone two MRI investigations (range 1–4). Because the Swiss Epilepsy Center has no MR scanner of its own, all images were acquired outside at other institutions and evaluated by external radiologists or neuroradiologists. In nearly half of the patients (19/40), PNH had been overlooked in at least one previous MRI, and in eight patients also in the high-resolution MRI subsequently used for the postprocessing presented herein. The lesions in these latter patients had been detected only by morphometric analysis, however, at that time only with visual inspection of the extension image, not by automated processing of the PNH image described herein. Most of the patients had focal epilepsy; only two patients did not have epilepsy but psychogenic seizures as documented by video–electroencephalography (EEG) monitoring. Demographic and clinical data are given in Table 1.

Table 1.   Demographic and clinical data of patients and controls
Age (mean ± 1 SD, range)32.8 ± 14.4, 4.3–65.832 ± 11.7, 16.5–74.5
Gender19 female, 21 male208 female, 192 male
Isolated PNH19
Multiple PNH21
Unilateral PNH, side23, right: 12, left: 11
Bilateral PNH17
Other epileptogenic lesions 8
PNH overlooked in the past19
PNH overlooked in the MRI used for postprocessing 8

The results of ROC analysis and automated PNH detection are presented in Fig. 3. The two diagrams labeled “ROC curve” and “Mirrored ROC curve” display the outcome for varying z-score thresholds. The AUC amounted to 0.96 (CI 0.92–1.00). The gray diagonals (“line of equality”) represent the results of a hypothetical random classifier with an AUC of 0.5. The optimal trade-off for the differentiation of true positives and false positives was determined by finding the cutoff point on the ROC curve closest to the corner of the diagram that represents the best possible performance (i.e., sensitivity and specificity of 100%). The z-score at this cutoff point (at the intersection of ROC curve and green diagonal) amounted to 6.6. Using this value as a threshold, PNH could be identified in 37 of 40 patients, corresponding to a sensitivity of 92.5% (CI 84.3–100.0%). The maximum z-scores found in the PNH images of these patients ranged between 6.6 and 71.5. False-positive findings with z-scores of 6.6–13.1 were observed in 34 of 400 controls, corresponding to a specificity of 91.5% (CI 88.8–94.2%). The vertical difference between the optimal cutoff point on the ROC curve and the line of equality derived from a random classifier represents the Youden index, another measure for the performance of the test (vertical blue line in Fig. 3). It maximizes the difference between true-positive and false-positive findings and was 0.84 in this study (a perfect test would have a Youden index of +1).

Figure 3.

Upper half: ROC curve and mirrored ROC curve for varying z-score thresholds (please cf. text for details). Lower half: Graphical plot showing the ratios of true-positive, true-negative, false-negative, and false-positive results for the optimal cutoff point on the ROC curve differentiating best between true positives and false positives.

Detection of previously overlooked PNH

PNH could be identified in 17 of the 19 patients in whom these malformations had not been detected in at least one previous MRI. Furthermore, PNH was found in all eight patients in whom the lesions had been overlooked, even in the high-resolution MRI acquired according to a dedicated epilepsy protocol, which was the basis for the MRI postprocessing shown here. To give an impression of these previously overlooked PNH cases Fig. 4A–F presents six examples of the most subtle lesions in the latter patient group.

Figure 4.

(AF) Detection of PNH lesions, which have been previously overlooked in the same high-resolution MRI that was used for postprocessing. Each row shows coronal slices of the T1 image, the coregistered T2 image, and the corresponding “extension image” of a single patient. (F) Example patient in whom MRI postprocessing not only detected the PNH but also highlighted abnormalities of the adjacent cortex. The enlarged cutouts in the third row show that parts of the cortical ribbon in this area resemble polymicrogyria (arrows). (GI) Failed detection of PNH in three patients in whom the maximum z-scores of the PNH images were found at the location of the PNH lesions (crosshairs) but did not exceed the predefined z-score threshold. (Jl) Examples of false-positive findings in three healthy controls: The upper row shows the T1 images, and the lower row the corresponding extension images. In two of these cases, automated PNH detection pointed to small areas of hypointensity at the tip of the frontal horn of the lateral ventricles, possibly due to transependymal migration of CSF into the extracellular space of adjacent periventricular white matter. This may resemble PNH in T1 images but usually appears hyperintense in relation to gray matter in T2 and FLAIR images.

Failed detection of PNH

There were three patients with previously diagnosed PNH that could not be detected by automated morphometric analysis. In these patients, the maximum z-scores found in the PNH images ranged between 2.1 and 5.6, and thus were below the predefined z-score threshold of 6.6. Figure 4G–H gives an impression of size and conspicuity of the PNH lesions in these false-negative cases.

Detection of associated cortex abnormalities

The region of interest around the lateral ventricles that defined the search space for ectopic GM in the PNH images sometimes also included parts of the adjacent or overlying cortex, especially in the midline region and in posterior brain areas. In a few patients this led to the situation that, apart from the detection of PNH lesions, morphometric analysis also pointed to possible abnormalities of the adjacent cortex, for example, sulci that appear to extend abnormally far into the white matter. Figure 4F demonstrates such an example, with PNH in the roofs of both lateral ventricles. The patient belongs to the eight cases in which PNH had been overlooked in the MRI used for postprocessing. However, morphometric analysis also highlighted the adjacent cortex in the midline region, and closer inspection reveals that parts of the cortical ribbon in this area have the appearance of polymicrogyria. This supports the hypothesis of an associated cortical malformation.

False-positive detection of PNH

There were false-positive findings in 34 (8.5%) of the 400 healthy controls. In most of these cases, the PNH detection algorithm highlighted small areas of periventricular hypointensity at the tip of the frontal horn of the lateral ventricles. These alterations are thought to result from transependymal migration of CSF into the extracellular space of adjacent periventricular white matter. Figure 4J–L presents three typical examples of false-positive findings.


In the present study, a novel method of automated morphometric MRI analysis for the detection of PNH has been evaluated. The need for additional diagnostic tools to detect PNH might appear surprising, because usually the visual detection of PNH in high-resolution MR images is considered to be unproblematic. PNH consists of nodules of ectopic GM and has the same signal intensity as mature GM in all MR sequences. Searching the predilection sites in the periventricular region allows the detection by visual assessment in many cases. However, some heterotopic nodules still escape radiologic detection (Meroni et al., 2009), and as also shown in this study, the percentage of PNH that is not detected visually is high. This is especially true for small or single PNH lesions. In clinical practice, the detection and diagnosis of PNH as the cause of epilepsy can be fundamental for further prognostic and therapeutic considerations (Aghakhani et al., 2005; Scherer et al., 2005). The presence of a circumscribed lesion on MRI supports the classification as focal epilepsy, an important factor in the choice of antiepileptic drugs. In addition, other therapeutic options such as epilepsy surgery or radiofrequency lesioning can be considered (Luders & Schuele, 2006; Stefan et al., 2007; Catenoix et al., 2008; Meroni et al., 2009; Guenot et al., 2011; Schmitt et al., 2011). Overall, the detection of PNH has clinical relevance and can be crucial for the patient.

The morphometric MRI analysis presented here combines an already known approach, that is, voxel-based morphometric analysis for the detection of ectopic GM (Kassubek et al., 2002; Wilke et al., 2003; Huppertz et al., 2005), with a new development that focuses the search for PNH on a region of interest around the lateral ventricles. This was done to increase sensitivity and specificity for automated detection of abnormally located GM especially in this area. It is achieved by using a binary mask based on the individual CSF space of the investigated patient, which (after smoothing/dilatation) includes a small band of brain parenchyma around the lateral ventricles. Ectopic GM in this area is detected by comparison of the individual GM distribution to a normal database and by searching for the maximum deviation from this normal database. As an additional criterion, a z-score threshold for differentiation of true-positive and false-positive findings is applied.

With a sensitivity of 92.5% and a specificity of 91.5%, the MRI postprocessing method presented herein seems to be a valuable aid in the detection of PNH. Nearly all patients in whom PNH had been overlooked at least once in the past (17 of 19) could be identified. In 20% of patients (8 of 40), PNH had not been detected, even in the high-resolution MRI, which was the basis for the analysis shown here, although these measurements were acquired according to a dedicated epilepsy protocol. They could all be identified by MRI postprocessing.

It is not unusual that PNH is combined with other cortical malformations, for example, polymicrogyria or other, more complex brain malformation (Leeflang et al., 2003; Wieck et al., 2005; Parrini et al., 2006). In line with this, the morphometric analysis also highlighted possible abnormalities of the cortex adjacent to the PNH in a few cases, for example, sulci in the midline region that extended abnormally far into the white matter and appeared to be polygyric (cf. Fig. 4F). Although this means that the intended confinement of the search for ectopic GM to the immediate periventricular region is not always perfect, the recognition of associated cortical malformations in the presence of PNH can be regarded as an additional advantage of the method.

Methodologic considerations

The prevalence of PNH in the general population is unknown and considered rather low (Tassi et al., 2005). In a large series of adult patients with epilepsy, a prevalence of about 2% has been reported (Raymond et al., 1994, 1995). In our database, PNH was found in about 1.5% of 3,719 patients. However, in studies of patients with cortical dysgenesis, PNH appeared to be the most common clinically encountered heterotopia and accounted for 15–20% of the patients (Raymond et al., 1994; Dubeau et al., 1995; Li et al., 1997; Battaglia & Granata, 2008). By choosing a proportion of 40 PNH patients versus 400 controls we tried to reproduce a prevalence that is amidst the values mentioned above and which might be representative of the population of patients with epilepsy in presurgical centers.

Our normal database for morphometric analysis comprised MRI data from five different MRI scanners of different field strengths. For this study is was necessary to analyze MRI measurements from different sources, even if no comparative data from the same scanner were available, and therefore a combined normal database was chosen. As already discussed before (Huppertz et al., 2008b), the theoretical disadvantage of not using a scanner-specific normal database is probably outweighed by the fact that our combined database is larger (currently 150 MRI control data sets) and thus a better model for the variability in the normal population.

The optimal z-score threshold for differentiation of true-positive and false-positive findings was based on the same population of PNH patients and controls for which the sensitivity and specificity of automated PNH detection were subsequently determined. Using subjects of the same sample both for definition of a classification parameter and for validation of the subsequent classification is usually discouraged, since this may lead to bias due to overfitting, that is, the results of validation may be better than for another sample comprising subjects not used for definition for the classification parameter. This study, however, was “exhaustive” in the sense that not only a sample but all available PNH patients in the predefined time period (i.e., 2006–2011) have been included. This reduces the risk of overfitting. Nonetheless, when implementing the proposed algorithm in other centers it seems reasonable to define anew the optimal z-score threshold on MRI data available at the respective site, to account for different scanner- and sequence-dependent signal inhomogeneities. Our study is only meant to show that automated PNH detection by morphometric analysis is principally possible and can achieve relatively high levels of sensitivity and specificity.

The presented method is completely automated and observer independent. As input, T1-weighted volume data sets are required, either from 1.5T or 3T MRI scanners. The script employs standard procedures of SPM5 and additional simple computations, mostly done by the image calculation tool of SPM5. The method is based on freely available software, except for the commercial MATLAB platform required for SPM. But there are also other freeware image-processing environments that are able to perform the key steps of this method. In summary, the method can easily be implemented in other centers.


The limitations of this method are reflected by the rate of false negatives and false positives. Three of 40 PNH patients could not be identified. Although the maximum z-scores of the PNH images were found at the location of the PNH, they did not exceed the required z-score threshold, probably because of spatial proximity to the caudate nucleus or thalamus. Reduction of the z-score threshold would reduce the rate of false negatives, but simultaneously increase the rate of false positives. For this study, a sensitivity of 100% (at a z-score threshold of 3.7) would have resulted in a specificity of only 27.5%. On the other hand, an increase of the z-score threshold would diminish the rate of false positives, but also the overall detection rate: a specificity of 100% (z-score threshold 13.1) correlates with a sensitivity of 50% in our study. That means, only half of the PNH would have been detected, probably a lower detection rate than by conventional visual assessment.

Among the false-positive controls there were many cases in which automated PNH detection pointed to small caps of hypointensity abutting on the poles of the lateral ventricles (cf. Fig. 4J–L). These findings resemble PNH in T1-weighted images but are thought to result from transependymal migration of CSF into adjacent periventricular white matter or to represent age-related alterations (Kertesz et al., 1988; Schmidt et al., 2011). In the clinical setting, it should be easy to differentiate them from true PNH lesions by help of their hyperintensity in T2- or fluid-attenuated inversion recovery (FLAIR) weighted images. In the future, the identification and rejection of these false positives could also be automatized by parallel processing of coregistered FLAIR images.

In the case of multiple or bilateral nodular heterotopia, the algorithm cannot identify all nodules but only those for which the z-scores exceed the given threshold (typically two to three nodules per patient). However, for clinical purposes it seems to be more important that the detection algorithm indicates the presence of PNH and directs the attention of the radiologist to them rather than to highlight all nodules.

Furthermore, the proposed method works best on brains that do not deviate too much from normal morphology. Due to the concept of using a predefined search space the detection of PNH is restricted to the space of the given mask derived from segmenting the individual lateral ventricles of the investigated patient. Therefore, the performance depends on the quality of delineating this search space. When there is a gross lesion or a malformation affecting the form and size of the individual lateral ventricles this could lead to both false-negative and false-positive findings. But it can be expected that such pathology is easily recognized and increases the attention and sensitivity for additional pathologic findings. PNH in an otherwise totally normal brain is more prone to be overlooked, and our method aims primarily at helping in this situation.


The method of morphometric MRI analysis for the detection of PNH presented here is the first approach that focuses an automated search for ectopic GM to the predilection sites of PNH, that is, the area around the lateral ventricles. Thereby, it increases the sensitivity for the detection of PNH, which according to our results is still at risk to be overlooked, even with high-resolution MRI. The automated image processing steps render the method objective and time-efficient, whereas the use of conventional T1 images without need for extra sequences is economic. Overall, the method facilitates the assessment of MRI data and appears to be a promising additional diagnostic tool in the evaluation of patients with epilepsy.


The development of the MRI postprocessing technique presented herein was kindly supported by the Swiss Epilepsy Foundation and the Swiss National Foundation. The work of Dr. B. Pascher in the field of MRI postprocessing at the Swiss Epilepsy Centre was financed by a fellowship of the Society for Neuropaediatrics (GNP, Gesellschaft für Neuropädiatrie).


None of the authors has any conflict of interest to disclose. We confirm that we have read the Journal’s position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.