Automated detection of abnormal changes in cortical thickness: A tool to help diagnosis in neocortical focal epilepsy

Transcription

1 Automated detection of abnormal changes in cortical thickness: A tool to help diagnosis in neocortical focal epilepsy 1. Introduction Epilepsy is a common neurological disorder, which affects about 1 % of the population in industrialized and even more in less developed countries. It is characterized by recurrent epileptic seizures, which are sudden excessive discharges of brain cells. Many factors can produce the epileptic state, for example head injuries, vascular damages, tumours or genetic factors; it is in fact more proper to make reference to the epilepsies. Seizures are classified into partial (focal or local), which start in a limited part of the brain, and generalized, where most of the brain is involved from the onset. In most patients, epilepsy can be treated with medication that typically aims at reducing the neuronal excitability, with efficiency depending on the type and causes of epilepsy. However, medication is inefficient in approximately 20 % of the patients: these patients are said to have refractory or pharmacoresistant epilepsy. When epilepsy is focal, i.e. when seizures start in a very limited part of the brain, a surgical procedure can be considered in order to remove the part of the brain responsible for the seizures. This removal (resection) of a part of the brain may appear like a drastic option, but one has to consider the severe handicap that arises from the epileptic condition. Also, an effort is made to avoid resecting regions that would lead to too severe postsurgical losses. Moreover, the resected region is often damaged, and the brain may have already compensated by involving other areas. Presurgical evaluation consists in combining many sources of information on the patient s epilepsy in order to define as precisely as possible the zone to be removed. The demonstration of a cortical lesion can significantly change the outcome of the surgery for epilepsy in neocortical cases. For small focal atrophies, cortical thickness is one of the most important parameters to be evaluated because it may point to possible destructive lesions or possible dysplasia. This work presents a semi-automated method for identifying and segmenting the lesions on T1-weighted MRI based on FreeSurface software analysis through the analysis of cortical thickness. The cortical thinning is usually regionally specific and the atrophy and its progress can, in some cases, reveal much about the evolution of the disease. The cortical mantle varies in thickness depending on the region of the cortex, with considerable variation between individual brains as well as between hemispheres of the same brain. In normal brains the cortical thickness varies between 1 and 4,5 mm, with an overall average of 2,5 mm. Being able to accurately estimate the thickness of the entire cortex of individual subjects, or group statistics for patient or control populations can translate in a tool for segment cortical lesions in necocortical focal epilepsy. In terms of software, FreeSurfer is a package developed at the A. Martinos Center for Biomedical Imaging at Harvard Medical School was used for the analysis of the MR-data. Freesurfer is a set of software tools for the study of cortical and subcortical anatomy. The software provides various analysis tools including: representation of the cortical surface between white and gray matter, representation of the pial surface, segmentation of white matter from the rest of the brain, skull stripping, B1 bias field correction, nonlinear registration of the

2 cortical surface of an individual with an sterotaxic atlas, labeling of regions of the cortical surface, statistical analysis of group morphometry differences, and labeling of subcortical brain structures. 2. Purpose and Participants In this present study we used Freesurfer to address the question of whether the thickness of cortical gray matter is reduced in patients with neocortical focal dysplasia and, if so, to determine the regional distribution of such thinning in each case. A morphometric study that yields measures of cortical gray matter thickness is applied so that homologous regions can be averaged and compared within and between subject groups. In order to do so, individual brains must be aligned, by registering to standardized volumetric space or by using computational matching strategies that align corresponding locations on the cortical surface. We present results from automated surface reconstruction, transformation, and highresolution intersubject alignment procedures for accurately measuring the thickness of the cerebral cortex across the entire brain as well as for generating crosssubject statistics in a coordinate system based on cortical anatomy in a cohort of patients with neocortical focal dysplasia. A group of two patients with medically refractory focal epilepsy and no lesion demonstrated in high resolution brain MRIs were selected. All underwent long term video-eeg monitoring (27 to 32 electrodes) to document the neurophysiological characteristics of their epilepsy as part of a comprehensive evaluation for epilepsy surgery. The MRIs consisted of high resolution (0.4x0.4x1.5 mm) volumetric T1 sequences including the whole brain. The control group consisted of 23 normal subjects aged years and submitted to a standard high resolution volumetric MRI, as part of a program of normalization of imaging studies for surgery of epilepsy at the Magnetic Ressonance Imaging center of Caselas. Thickness maps from the 23 normal control subjects were averaged using Freesurfer s highresolution surface-based averaging techniques and compared with the thickness measurements from the thickness maps of each patient subjects individually. Mean cortical thickness and variance of mean were calculated at each location. The statistical maps were generated using a random effects model with 1 degree of freedom for each subject to generate a t-test for each cortical location. 3. Results 3.1. Control Group After all the individual images were preprocessed by FreeSurfer, the individual thickness estimates from the results of this analysis were combined across the 23 normal by using the high-resolution, surface-based averaging technique that aligns cortical folding patterns. A 7 mm fwhm Gaussian blurring kernel was used for smoothing of the cortical thickness measures surface. Thickness measures were mapped to the inflated surface of each participant's brain reconstruction, allowing visualization of data across the entire cortical surface (i.e.gyri and sulci) without being obscured by cortical folding.

3 Left Hemisphere Right Hemisphere Left Hemisphere Right Hemisphere Figure 1 Inflated surface with thickness map of the average control group. The inspection of the control group allowed recognition of several features: The primary sensory areas were easily identified as extended areas of reduced cortical thickness The motor cortex is also easily identified as an extended area of increased cortical thickness Most of the cortex shows a pattern of patches of decreased and increased thickness, corresponding respectively to the depth of sulci and the top of gyri A light asymmetry between the two hemispheres: the right one appears to have, in general, an average cortical thickness inferior to the left one. This result is consistent with published findings. An illustration of the variability of these results across the cortex is given in Figure 2, which shows the spatial distribution of the cross-subject standard deviations of the thickness measurements. As can be seen, the measurements are quite consistent across subjects, with a standard deviation of less than 0.5 mm over much of the cortex (grey areas). No major asymmetries were apparent between hemispheres. Nevertheless, notice that the majority of the variance is localized in areas of higher cortical thickness. Left Hemisphere Right Hemisphere Left Hemisphere Right Hemisphere Figure 2 - Map of the standard deviations of the thickness measurements across 23 subjects

4 3.2. Overall picture of the approach The principle underlying our approach consists in finding any abnormality regarding the cortical thickness of the patient. Further analysis is required to successfully differentiate lesional tissue from healthy cortex because of the poor difference in terms of grey level. For example, in heterogeneous lesions, the white/grey matter interface can be blurry, which may lead to a bad white and pial surface generation. A simple visual inspection of cortical thickness maps is not enough to delineate the abnormal area. For that purpose, the method we use is made up of four different separate steps: 1. Visual inspection of the thickness patterns in order to identify areas of asymmetry between hemispheres or through changes in the normal cortical patch pattern 2. Analyze the detected abnormal areas in the 3D MRI, looking for eventual problems during the surface reconstruction 3. Define Regions of Interest in the surface in order to perform an analysis in terms of thickness values, variability differences, sulcal or gyral location, etc 4. Define the appropriate statistic for the case 3.3. Patient 1 Visual inspection of the thickness patterns The processing of the MRIs from both the patient and the control group produced a final a final colour representation of the cortical thickness of each hemisphere. From the inspection of the patient s maps and its interhemispheric comparison, an abnormal pattern of reduced cortical thickness was spotted in the right occipital lobe. The abnormal area was traced in order to compare it with the control group. This step is of extreme relevance because it is indeed very hard to visually compare a single subject inflated surface with an average group due to their surface s topological differences. Regional comparisons between the patient and the normal group confirm that the former has a cortical thinning within the region of interest: the patient presents a single colour instead of the patchy pattern found in the average normal subject. Notice that not all the region inside the defined ROI is necessarily an atrophy. In fact, a great part of that light blue area in the patient s map corresponds to an also light blue area in the group s one. In that, the defined region must be shrunk and better determined in order to precisely delineate the lesion s area.

5 Left Hemisphere Right Hemisphere Figure 3 Cortical Thickness map of Patient 1. vs Control Group From interhemispheric comparison a light blue abnormal pattern is found in only one of the hemispheres. Figure 4 Abnormal cortical thickness pattern in the volumetric MRI. The previously defined label (in yellow and purple in the cortical thickness maps) is her shown as the green area. These magnetic resonance images seem to confirm the atrophy. Define Regions of Interest Delimitating the lesion is anything but linear inasmuch as cortical thickness varies with geometry, more specifically, gyral regions are thicker than sulcal ones. In order to evaluate if this fact has any effect in our defined region, the location of the potential atrophy was studied and the region of interest divided in two different parts: the gyral and the sucal one. The label s sulcal region is identical in both hemispheres in terms of cortical thickness pattern. This pattern is also very similar to the average normal group one. Therefore, this region can be removed of the region of interest. Nevertheless, this simplification is not enough yet. Actually, it is clear from the plot of the mean cortical thickness across subjects that this whole area does not define an atrophy. The patient has a mean cortical thickness within the range of the control group mean cortical thickness. This means that the lesion has not been well located yet. Comparing cortical thickness pattern within the new defined gyral label with the control group, a much smaller label is defined (Figure 5). This label contains only the vertexes where the patient has inferior cortical thickness values than most of the normal subjects. These vertexes were manually found by visual inspection of the mean cortical thickness distribution. With this atrophic region delineated, the real boundaries of the lesion can be drawn by carefully inflating the label. Transferring to the 3D MRI volume, it is possible to enlarge the initial area and identify the whole region that strikes us as being part of the lesion. From this inspection, a final label is found with acceptable mean cortical thickness values.

6 Figure 5 and 6 Atrophy label in the inflated surface and Mean cortical thickness values within the new found label. The former label, delimitated in purple, between gyrus and sulci. In yellow is the new defined gyral label. On the right, patient s is below most of the normal subjects. This fact associated with a low standard deviation value seems to confirm the suspicions of an atrophic lesion in the define area. Figure 7 and 8 (Figure 7) Significance map. The atrophy is delineated in purple. Notice the high significance values (red) within the label area. (Figure 8) Abnormal areas (in green) in the volumetric MRI. The coloured area presented here is the same area delimitated by label in the inflated surface. The pial (red) and white (yellow) surfaces are also shown. From visual inspection it is clear this area is a lesion with reduced cortical thickness Patient 2 Visual inspection of the thickness patterns Comparing the patient s individual thickness estimates with the control group, two different areas were spotted as potential abnormal thickness patterns. The first one is a small area of reduced cortical thickness in the left occipital lobe (Figure 9). The second one is a slightly bigger area, also in the left hemisphere, in the interior occipital lobe (Figure 10). Figure 9 Cortical Thickness Maps in inflated surface of the control group (left) and patient 2 (right). From visual inspection, a darker blue pattern identified in the control group does not seem to appear in the patient s map

7 Figure 10 - Cortical Thickness Maps in inflated surface of the control group (left) and patient 2 (right). From visual inspection, a light blue pattern is identified in the control group does not seem to appear in the patient s map In order to confirm these suspicions, the spotted abnormal pattern is marked out in both regions. The first region was traced on the control group and transposed to the patient s inflated surface. After tracing the apparently different pattern on the control group thickness map, it is possible to carry the traced region over to Patient s 2 map. This darker blue area found on the control group actually also corresponded to a darker blue area on the patient s thickness map. In other words, there may not be a difference between patterns. From mean cortical thickness estimation no major discrepancy was found between the group and the patient s values. As the label has already a reduced size, there is no need to continue the analysis on this area. The difference of pattern identified in the cortical thickness was not a lesion. When it comes to the second potential lesion, the significance map (Figure 11) seems to be consistent with the suspicions. The patient presents a significant inferior cortical thickness region within the delimitated label (red spot). Figure 11 and 12 Significance map (left) and mean cortical thickness distribution within the delimitated label (right). The atrophy is delineated in purple. Notice the high significance values (red) within the label area. This is validated by the mean cortical thickness plot. Within the label, patient 2 has lower cortical thickness values than every other normal subject. However, notice the high standard deviation associated to this measure. Analysis of the detected abnormal areas in the 3D MRI Problems with respect to ROI analysis arise because of the manual way in which the ROIs are placed. In the first step of this analysis the ROI is placed in a particular region, where a cortical

8 thickness abnormality seemed to be present. In order to confirm the location of the potentially found lesion, visual inspection and analysis of its placement in the 3D MRI volume is required. As the first region has already been excluded of our analysis, we will only focus on the second. This region of interest, despite the unusual pattern found by comparison with the normal group and the low mean cortical thickness value, had no visible lesion in the MRI. This label can also be excluded of the analysis. Although none of the former identified labels corresponded to a real lesion, another potential atrophy was found by visual inspection of the MRI volume. Analyzing the MRI an ambiguous region is identified as lesional. There is a clear heterogeneous lesion in the interior face of the left occipital lobe. Cortical thickness changes were not identified in the first step because of the blurred grey/white matter interface. FreeSurfer s automated grey/white matter segmentation seems to have failed and the generated surfaces are erroneous. Thus, it is not possible to continue an analysis strictly based on cortical thickness maps comparisons. As it is easily identified, the lesional area was delineated in the MRI. In order to define a more reliable region, we analyzed the cortical thickness maps once more. From interhemispherical visual comparison a smaller region of interest within the previously defined lesional area is delimitated. Define Regions of Interest Comparing with the control group, no abnormal pattern is identified within the label. However, if compared with any normal subject MRI, the patient s label clearly delimitates an anomaly. In the significance maps, there appear several regions with positive and negative values of significance, however none of these are related to the label This result comes as no surprise, since the pial and white surfaces reconstruction seems to have failed. Patient s mean cortical thickness values also appear normal compared with the control group (Figure 14). Figure 13 and 14 New label definition and Mean cortical thickness distribution within the potential lesion region. From MRI analysis a new label is defined. Cortical thickness patterns do not seem to differ when comparing patient 2 with control group thickness map. The patient s value also seems normal when compared with the control group. Nevertheless, notice that the standard deviation is extremely high.

9 Parcelation In order to confirm the weak boundary issue, the standard deviation was computed all across the cortex. The basic idea here is to compute the dispersion associated to thickness measurements in order to determine the liability of the white and pial surfaces determination in the lesional area. FreeSurfer offers a completely automated cortex segmentation tool. It segments the whole cortex in several anatomical regions. FreeSurfer can also estimate thickness measures (average cortical thickness ± standard deviation) for each label. There are three areas stand out as having higher levels of dispersion. These regions are all located near the lesion area confirming the suspicion of erroneous surface determination. In cases of heterogeneous lesions such as this, this method can be very useful to determine a rough location of the lesion since the algorithm seems to have troubles defining these areas. 4. Discussion In this section, we proposed an approach to improve determination of the lesion location. The method relies on MRI-feature knowledge and a surface-based analysis of cortical thickness maps. The method requires a great amount of user interaction. Most of the analysis is based on visual inspection and evaluation. This approach can be improved if instead of performing the parcelation in the end of the analysis, we do it in the beginning. Evaluation of the parcelation results can give us a hint of where the lesion is located. This would spare the user of looking through the whole cortex in search of abnormal patterns. In that, our final proposal is that the analysis is conducted according to the flowchart below. 5. General discussion and conclusions In this manuscript we proposed and evaluated a novel semi-automated workflow for locating small cortical lesions in patients with focal neocortical epilepsy on high-resolution MRI. The precise delineation is crucial for clinical diagnosis and surgical planning, but their MRI features make this task challenging. The blurred grey-white matter interface and the absence of evident boundaries between dysplasic tissues and healthy cortex may lead to misdiagnoses; moreover, because the cortex is only a few millimetres thick and is highly irregular, it is also very difficult to study it directly by conventional MRI evaluation. In addition, most of the cortex lies hidden in sulci and is not visible in direct visual inspection at the time of the surgery. To figure out these difficulties and successfully find the lesions, we made use of software package Freesurfer and many of its tools. The procedures of surface cortical reconstruction have already been proved reliable in several studies when using high quality datasets. This surface based method improves visualization when compared with conventional volumetric representation of

10 the brain. The procedure of cortical surface inflation allows a complete visualization of the cortex and eventual abnormalities along the surface are more easily identified and evaluated. Furthermore, topological relations are preserved, allowing simultaneous visualization of cortical point in the inflated/pial/white surface and in the volumetric image, improving the ability to evaluate the significance of the findings. Measuring thickness of the cortex, as well as the segmenting it for parcelation, are completely automatic procedures, which gives a certain reliability to the method. This type of analysis provides a certain level of automatization, making the whole process of lesion search more precise and less prone to visual errors. The visual inspection allows not only detection of local spots of increased/decreased thickness, but also differences in the normal spatial pattern along the cortex. Additional strategies, as sulcal/gyral location and definition of statistical analysis within certain regions of interest were implemented in order to improve and assure lesion coverage. The co-registration procedure employed here makes use of the pattern of folding across the entire cortical surface, and is a fully automated process done by FreeSurfer. Once the initial alignment of all the cortical surfaces has been accomplished, they are averaged in order to generate a probabilistic atlas as the target for the final registration procedure. This is an extremely powerful tool for group analysis and creation of a control group database. Our results suggest that the high-resolution volumetric MRI datasets obtained in these patients contain more information than the one that is extracted by the traditional visual analysis, and that further computer analysis is able to demonstrate previously unknown cortical lesions. The obtained results encourage us to consider this technique as a useful tool for visual diagnosis. This semi-automated method can constitute an objective criterion and may unveil subtle lesional areas that could have been overlooked by the expert. Moreover, with improvements and further validation, this method may be part of a pre-surgical protocol defining the extent of the tissue to resect.