This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/authorsrights
Author's personal copy Epilepsy Research (2014) 108, 481—490
journal homepage: www.elsevier.com/locate/epilepsyres
Visualizing Meyer’s loop: A comparison of deterministic and probabilistic tractography Ylva Lilja a,e,∗, Maria Ljungberg b, Göran Starck b, Kristina Malmgren c,e, Bertil Rydenhag d,e, Daniel T. Nilsson d,e a
Ear, Nose and Throat Clinic, Sahlgrenska University Hospital, Gröna Stråket 5, 41345 Gothenburg, Sweden Department of Medical Physics and Medical Engineering, Sahlgrenska University Hospital, 41345 Gothenburg, Sweden c Department of Neurology, Sahlgrenska University Hospital, 41345 Gothenburg, Sweden d Department of Neurosurgery, Sahlgrenska University Hospital, 41345 Gothenburg, Sweden e Institute of Neuroscience and Physiology, Department of Clinical Neuroscience and Rehabilitation, University of Gothenburg, Per Dubbsgatan 14, 41345 Gothenburg, Sweden b
Received 22 May 2013; received in revised form 29 October 2013; accepted 14 January 2014 Available online 2 February 2014
KEYWORDS Optic radiation; Meyer’s loop; Diffusion tensor imaging; Tractography; Temporal lobe; Epilepsy surgery
Summary Background: Diffusion tensor tractography of the anterior extent of the optic radiation — Meyer’s loop — prior to temporal lobe resection (TLR) may reduce the risk for postoperative visual field defect. Currently there is no standardized way to perform tractography. Objective: To visualize Meyer’s loop using deterministic (DTG) and probabilistic tractography (PTG) at different probability levels, with the primary aim to explore possible differences between methods, and the secondary aim to explore anatomical accuracy. Methods: Twenty-three diffusion tensor imaging exams (11 controls and 7 TLR-patients, preand post-surgical) were analyzed using DTG and PTG thresholded at probability levels 0.2%, 0.5%, 1%, 5% and 10%. The distance from the tip of the temporal lobe to the anterior limit of Meyer’s loop (TP—ML) was measured in 46 optic radiations. Differences in TP—ML between the methods were compared. Results of the control group were compared to dissection studies and to a histological atlas.
Abbreviations: TLR, temporal lobe resection; DTG, deterministic tractography; PTG, probabilistic tractography; TP—ML, distance from tip of temporal lobe to tip of Meyer’s loop; FA, fractional anisotropy; ROI, region of interest; LGN, lateral geniculate nucleus. ∗ Corresponding author at: Ear, Nose and Throat Clinic, Sahlgrenska University Hospital, Gröna Stråket 5, 41345 Göteborg, Sweden. Tel.: +46 31 3429590; fax: +46 31 829615. E-mail addresses:
[email protected],
[email protected] (Y. Lilja),
[email protected] (M. Ljungberg),
[email protected] (G. Starck),
[email protected] (K. Malmgren),
[email protected] (B. Rydenhag),
[email protected] (D.T. Nilsson). 0920-1211/$ — see front matter © 2014 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.eplepsyres.2014.01.017
Author's personal copy 482
Y. Lilja et al. Results: For controls and patients together, there were statistically significant differences (p < 0.01) for TP—ML between all methods thresholded at PTG ≤1% compared to all methods thresholded at PTG ≥5% and DTG. There were no statistically significant differences between PTG 0.2%, 0.5% and 1% or between PTG 5%, 10% and DTG. For the control group, PTG ≤1% showed a closer match to dissection studies and PTG 1% showed the best match to histological tracings of Meyer’s loop. Conclusions: Choice of tractography method affected the visualized location of Meyer’s loop significantly in a heterogeneous, clinically relevant study group. For the controls, PTG at probability levels ≤1% was a closer match to dissection studies. To determine the anterior extent of Meyer’s loop, PTG is superior to DTG and the probability level of PTG matters. © 2014 Elsevier B.V. All rights reserved.
Introduction Temporal lobe resection for therapy-resistant temporal lobe epilepsy is a well-established, safe procedure with few complications (Rydenhag and Silander, 2001; Wiebe et al., 2001). The most common neurological deficit after this procedure, affecting 50—90% of the patients, is a contralateral partial or complete upper quadrantanopia, caused by damage to the most anterior part of the optic radiation, Meyer’s loop (Egan et al., 2000; Guenot et al., 1999; Hughes et al., 1999; Jensen and Seedorf, 1976; Marino and Rasmussen, 1968; Nilsson et al., 2004). Dissection studies have shown a considerable interindividual variation in the anterior extent of Meyer’s loop (Chowdhury and Khan, 2010; Ebeling and Reulen, 1988; Peuskens et al., 2004; Rubino et al., 2005). Diffusion tensor based tractography has important clinical applications in neurosurgery including preoperative planning and intraoperative delineation of eloquent white matter tracts (Chen et al., 2007; Lee et al., 2004; Lo et al., 2007; Melhem et al., 2002; Mori and Zhang, 2006; Nguyen et al., 2005; Nimsky et al., 2005, 2006; Okada et al., 2006). Tractography of the optic radiation may predict the probability for a post-operative visual field defect after temporal lobe resection and reduce the risk of intraoperative injury (Chen et al., 2009; Thudium et al., 2010; Winston et al., 2012; Yogarajah et al., 2009). There is currently no standardized way to perform tractography of the optic radiation. The curving course of its anterior-most portion, Meyer’s loop, its close relation to CSF-filled spaces and its thin sheet-like appearance are properties that make the optic radiation a challenging structure to delineate using tractography (Sherbondy et al., 2008). Diffusion tensor imaging detects the amount and the direction of the water diffusion in each voxel. From this information the main direction of water diffusion and the size of the main water diffusion vector can be determined (Basser, 1995). A frequently used measure of the directionality of water diffusion is the fractional anisotropy (FA), where 0 is complete isotropy as in free water and 1 is complete anisotropy, as in unidirectional water diffusion. The main direction of water diffusion has been shown to reflect the direction of the regional white matter tracts and the FA value reflect the anisotropy of the tissue, where high FA values are found in highly organized white matter (e.g., corpus callosum) and low FA in tissue with free movement of water (e.g., CSF) (Mori and Zhang, 2006). By connecting voxels based on their anisotropy and their principal diffusion direction, images of the major white matter pathways can
be constructed. This is referred to as fiber tracking or ‘tractography’ (Basser, 1995; Conturo et al., 1999; Jones, 2008; Mori and van Zijl, 2002). Tractography algorithms can be classified into two types: deterministic and probabilistic. Deterministic, or streamline, tractography techniques assume that there is one main diffusion direction in each voxel and by defining a starting point and applying an algorithm linking voxels with similar diffusion directions a tract is produced (Jones, 2008). The technique can be refined by applying constraining criteria, including criteria for the fractional anisotropy, angle of deviation and multiple regions of interest where the tract must pass, a technique known as ‘‘virtual fiber dissection’’ (Catani et al., 2002). The advantages of DTG are relatively simple calculations and fast results with a clear delineation of fiber tracts, and the main downsides are operator-dependency, difficulties resolving curving, crossing or ‘‘kissing’’ tracts and that there is no indication of the confidence that one can assign to a reconstructed trajectory (Jones, 2008). In the most common, commercially available neuronavigation systems, deterministic tractography algorithms are used. In contrast to deterministic techniques, probabilistic tractography traces a large number of (typically >5000) possible pathways from a set starting point. At each step in the process of tracking a connection between voxels several different possible diffusion directions are considered (Jones, 2008). The result is a probability distribution of connections and by selecting an appropriate threshold, below which connections are discarded as unlikely, tracts can be outlined. The selection of this probability threshold has been arbitrary in most studies, even though Clatworthy et al. recently tried to ‘‘standardize’’ threshold selection by comparing tractography results with previously collected histological data from a different group of subjects (Clatworthy et al., 2010). Additional criteria (e.g., to only include fibers to occipital cortex) are often used to constrain the number of connections. The advantages of PTG are improved sensitivity also for curving fibers and improved resolution of crossing/kissing tracts compared to DTG (Jones, 2008). The trade-off is long computation times, decreased specificity resulting in ‘‘false-positives’’ and a less intuitive visualization of data as a ‘‘probability of a tract’’. Existing reports on tractography of the optic radiation have used either a deterministic (Chen et al., 2009; Nilsson et al., 2007; Taoka et al., 2008; Thudium et al., 2010; Yamamoto et al., 2005) or a probabilistic algorithm (Powell et al., 2005; Sherbondy et al., 2008; Winston et al., 2011,
Author's personal copy A Comparison of Deterministic and Probabilistic Tractography 2012; Yogarajah et al., 2009). Advocators of the probabilistic techniques have argued that it produces more anatomically accurate data, which has been supported by the theoretical framework for tractography algorithms (Behrens et al., 2003; Sherbondy et al., 2008). Supporters of the deterministic techniques have claimed these to produce accurate delineations of the optic radiation in a shorter time (Catani et al., 2003; Mori and van Zijl, 2002). Although differences between the methods have been seen previously when comparing results from different studies, no one study has directly compared DTG and PTG of Meyer’s loop, nor different probability levels of PTG of the optic radiation, using results from the same MRI protocol and a single set of DTI scans. The primary aim of this study was to directly compare DTG and PTG at different probability levels, to investigate possible differences in the delineation of Meyer’s loop. For this purpose, a heterogeneous study group with a mix of controls and patients was chosen, in an attempt to reflect the clinical reality. The secondary aim was to identify the method with the most anatomically accurate delineation, by comparing results of the control group to dissection studies (Chowdhury and Khan, 2010; Ebeling and Reulen, 1988; Peuskens et al., 2004; Rubino et al., 2005) and to a histological atlas (Burgel et al., 1999; Clatworthy et al., 2010; Mazziotta et al., 2001).
Subjects and methods DTI was performed in eleven controls (mean age 34 years, range 23—62 years) without neurological or psychiatric disease and in seven patients (mean age 36 years, range 15—58 years) with refractory temporal lobe epilepsy before (in five patients) and after temporal lobe resection (seven patients). The study was approved by the regional ethical board of the Gothenburg University, and informed written consent was obtained from all subjects.
MRI acquisition MRI was performed on a Philips Gyroscan Intera 1.5 T equipped with software release 9 with research software functionality HARDI (High Angular Resolution Diffusion Imaging, Philips, Eindhoven, The Netherlands) in 19 of the subjects. In four subjects a Philips Achieva 1.5 T scanner with software release 2.6 was used. A six-element SENSE head coil was used for the Intera and an eightelement head coil for the Achieva (Philips, Eindhoven, The Netherlands). The subject’s head was firmly supported with cushions. DTI of the whole brain was performed with the following parameters: single shot spin echo planar imaging with TE = 69 ms, TR = 10,190 ms, axial slices, SENSE factor 3.2, number of signals averaged 6, half scan factor 0.712, isotropic 2.2 mm × 2.2 mm × 2.2 mm voxels, b = 0 s/mm2 plus 15 diffusion-sensitizing gradient directions (b = 800 s/mm2 ) and phase encoding (bandwidth 33.8 Hz per pixel in the Intera and 36.4 Hz per pixel in the Achieva) in the anterior—posterior direction. Total scan time was 16 min. A T1-weighted scan (3D T1-TFE) was performed as an anatomical reference using identical orientation and slice thickness as the DTI scan. Reconstructed pixel size was 1.9 mm × 1.9 mm for both DTI and T1-weighted images.
483
Deterministic tractography DTI data was transferred to a workstation (8 GB RAM, 2.33 GHz Intel Xeon, Windows 64 bit). The ‘‘FiberTrak’’ package included in the Extended MR Workspace (EWS) R 2.6.3 (Philips Medical Systems Veenpluis 4-6, 5684 Best The Netherlands) was used. This fiber tracking software is based on the ‘‘fiber assignment by continuous tracking’’ (FACT) algorithm (Mori and van Zijl, 2002; Stieltjes et al., 2001). To minimize artifacts caused by movement and eddy currents, coregistration to b = 0 scans was done using the diffusion registration package in the EWS. Fractional anisotropy maps were generated in the EWS. Settings for the tractography were the following: FA > 0.25, max angle change 90◦ , min fiber length 30 mm. Multiple regions of interests (ROIs) were selected to ‘‘dissect’’ the optic radiation in a similar way as described previously (Chen et al., 2009; Conturo et al., 1999; Nilsson et al., 2007; Stieltjes et al., 2001). ROIs were placed as follows: ROI 1, all temporal lobe white matter (sagittal stratum) in the coronal plane at the level of superior colliculi; ROI 2, coronal plane, large ROI in occipital subcortical white matter; ROI 3, sagittal plane, entire midline ROI (exclude algorithm) to exclude fiber crossing midline. After this step the first tractography was calculated. By defining additional ROIs the optic radiation fibers were outlined in the following way: ROI 4, coronal plane, anterior temporal ROI, exclusion of anterior temporal fibers (exclude algorithm); ROI 5, sagittal plan: outline fibers entering in the region of posterior thalamus (lateral geniculate nucleus); ROI 6, axial plane, ROI over corticospinal tract fibers (exclude); ROI 7 and 8, if needed obviously aberrant fibers (to cerebellum, fornix etc.) were excluded. Images of the resulting tract in the coronal, axial and sagittal plane were saved. The distance between the tip of temporal lobe to the anterior edge of Meyer’s loop (TP—ML) was measured (Fig. 1). Time for diffusion registration, creation of FA maps and tensor calculation was approximately 10 min. Tractography of the optic radiation took 20—30 min per side.
Figure 1 ‘‘TP—ML’’: axial slice (b = 0) of probabilistic tractography of Meyer’s loop. Distance from the temporal pole to Meyer’s loop (TP—ML, in blue) is shown.
Author's personal copy 484
Probabilistic tractography DTI data was transferred to a workstation (6 GB RAM with a 2.67 GHz Intel Xeon quad core CPU, GNU/Linux: Ubuntu 9.04) where all subsequent calculations were done. The FMRIB’s Diffusion Toolbox (FDT) which is a part of the FMRIB software library (FSL) package (University of Oxford) was used for the post-processing (Behrens et al., 2003; Smith et al., 2004; Woolrich et al., 2009). After conversion of the DICOM data to the required Nifti-format, coregistration was done to b = 0 to correct for eddy current and movement artifacts. Then a whole-brain tractography was made using the FDT tool in FSL (Behrens et al., 2003), allowing modeling of crossing fibers within each voxel (Behrens et al., 2007), which took approximately eight hours to run. Angular threshold was set to allow a maximum 90degree angle deviation between voxels; default angular threshold in FSL is 80◦ , which was considered too restrictive in the case of Meyer’s loop, due to its known curving course. ROIs for tractography were constructed in FSLView (Smith et al., 2004). Initially a ROI was placed in the optic tract and the resulting tracts were used to help locating the lateral geniculate nucleus. A second ROI was placed in the lateral geniculate nucleus (LGN) (lateral to the transition between the cerebral peduncle and the posterior limb of the internal capsule) including the voxels lateral and anterior to the LGN. To ensure symmetrical ROI placement, ROIs were placed bilaterally in the same session. An additional ROI was placed in the subcortical occipital lobe. Finally, to exclude aberrant fibers anteriorly a temporal exclusion ROI was placed, 15 mm anterior to the temporal horn of the lateral ventricle. Previous work has shown that Meyer’s loop ends posterior to this (Burgel et al., 1999; Ebeling and Reulen, 1988; Kier et al., 2004). A similar approach to exclude anterior erroneous fibers was used in a previous PTG study (Yogarajah et al., 2009). The resulting tractography was thresholded at probability levels 10/5000 (1%), 25/5000 (0.2%), 50/5000 (1%), 250/5000 (5%) and at 500/5000 (10%). TP—ML distances were measured for the respective tractographies of the optic radiations (Fig. 1). ROI selection and tractography of the optic radiation took between 20 and 30 min for each side. To assess the reproducibility of the method, the ROI placement and tractography was carried out by two different raters independently as well as twice by one rater (the latter with a time separation of at least one month), for all the optic radiations. Inter- and intra-rater variability were achieved by comparing TP—ML distances. Probabilistic tractography of the optic radiation from the 11 normal controls were registered to MNI (Montreal Neurological Institute) space and overlaid on the same MNI brain and directly compared to the optic radiation in the Juelich atlas (Eickhoff et al., 2005). This histological atlas is composed of data from tracings of the optic radiation from 10 specimen aged 37—85 years combined in a group ‘‘probability map’’ of the optic radiation and coregistered to the MNI brain (Burgel et al., 1999; Eickhoff et al., 2005). These data are freely available online as part of the FSL software (Anatomy Toolbox, version 1.5, Jülich Research Institute, Germany). Tractography was carried out in a
Y. Lilja et al. standardized way by placing a seed ROI in the LGN and a waypoint ROI in the primary visual cortex, both derived from the Juelich atlas included in FSL package and identical in all subjects.
Statistical analysis Differences in TP—ML distances of tractographies from PTG at the different probability levels (0.2%, 0.5%, 1%, 5% and 10%) and of DTG were compared in an analysis of variance using mixed covariance pattern model, adjusting for dependences within individuals (the pre and post surgery scans of the 5 patients). Multiplicity was adjusted for using the Tukey—Kramer method. The procedure was repeated for all subjects together as well as for the control and patient group separately. Inter- and intra-rater variability were analyzed with distribution of differences between the raters’ measurements, limits of agreement and Intraclass Correlation Coefficient (ICC) (Shrout and Fleiss, 1979). In addition intra-rater variability was also analyzed with intra individual standard deviation. Systematic differences between the two raters and between the test—retest measurements were analyzed with Wilcoxon Signed Rank test.
Results Comparison of methods For the whole study group there were statistically significant differences (p < 0.01) between all methods thresholded at PTG ≤1% compared to all methods thresholded at PTG ≥5% and DTG. There were no statistically significant differences between the methods PTG 0.2%, 0.5% and 1% or between the methods PTG 5%, 10% and DTG (Table 1). The corresponding results for the control group and the patient group separately can be seen in Fig. 2. For the control group there were statistically significant differences (p < 0.01) between all methods thresholded at PTG ≤1% compared to all methods thresholded at PTG ≥5% and DTG. The same was true for the patient group, except for one comparison: there was no statistically significant difference between PTG 1% and DTG (p = 0.07). There were no statistically significant differences between the methods PTG 0.2%, 0.5% and 1% or between the methods PTG 5%, 10% and DTG, for neither the control group nor the patient group. The inter- and intra-rater variability tests, based on TP—ML measures of the 23 scans, showed good to excellent agreement (Fleiss, 1986) (ICC 0.6—0.8 inter-rater, ICC 0.7—0.9 intra-rater).
Anatomical accuracy — comparison to dissection data TP—ML measures for the 11 controls were compared to investigate anatomical accuracy (Table 2). There was no significant correlation between age and TP—ML for any of the tractography methods (Pearson’s R −0.17 to 0.05). PTG ≤1% showed a closer match to dissection studies than PTG ≥5% and DTG (Table 3).
Author's personal copy
Difference between the 6 models, using mixed covariance pattern model with adjustment for dependence within individuals and adjustment for multiplicity with the Tukey—Kramer method. The upper right part of the table shows p-values and the lower left part the difference estimates with 95% confidence intervals within brackets (lower; upper). TP—ML, distance between temporal pole and anterior border of Meyer’s loop. PTG, probabilistic tractography. DTG, deterministic tractography.
— −1.78 −2.93 −9.38 −10.66 −9.92
(−4.59; 1.04) (−5.75; −0.12) (−12.19; −6.56) (−13.47; −7.84) (−12.73; −7.11)
0.45 — −1.16 −7.60 −8.88 −8.14
(−3.97; 1.66) (−10.41; −4.79) (−11.69; −6.07) (−10.96; −5.33)
