The utility of customised tissue probability maps and templates for patients with idiopathic normal pressure hydrocephalus: a computational anatomy toolbox (CAT12) study

Abstract

Background

Disproportionately enlarged subarachnoid space hydrocephalus (DESH) is one of the neuroradiological characteristics of idiopathic normal pressure hydrocephalus (iNPH), which makes statistical analyses of brain images difficult. This study aimed to develop and validate methods of accurate brain segmentation and spatial normalisation in patients with DESH by using the Computational Anatomy Toolbox (CAT12).

Methods

Two hundred ninety-eight iNPH patients with DESH and 25 healthy controls (HCs) who underwent cranial MRI were enrolled in this study. We selected the structural images of 169 patients to create customised tissue probability maps and diffeomorphic anatomical registration through exponentiated Lie algebra (DARTEL) templates for patients with DESH (DESH-TPM and DESH-Template). The structural images of 38 other patients were used to evaluate the validity of the DESH-TPM and DESH-Template. DESH-TPM and DESH-Template were created using the 114 well-segmented images after the segmentation processing of CAT12. In the validation study, we compared the accuracy of brain segmentation and spatial normalisation among three conditions: customised condition, applying DESH-TPM and DESH-Template to CAT12 and patient images; standard condition, applying the default setting of CAT12 to patient images; and reference condition, applying the default setting of CAT12 to HC images.

Results

In the validation study, we identified three error types during segmentation. (1) The proportions of misidentifying the dura and/or extradural structures as brain structures in the customised, standard, and reference conditions were 10.5%, 44.7%, and 13.6%, respectively; (2) the failure rates of white matter hypointensity (WMH) cancellation in the customised, standard, and reference conditions were 18.4%, 44.7%, and 0%, respectively; and (3) the proportions of cerebrospinal fluid (CSF)-image deficits in the customised, standard, and reference conditions were 97.4%, 84.2%, and 28%, respectively. The spatial normalisation accuracy of grey and white matter images in the customised condition was the highest among the three conditions, especially in terms of superior convexity.

Conclusions

Applying the combination of the DESH-TPM and DESH-Template to CAT12 could improve the accuracy of grey and white matter segmentation and spatial normalisation in patients with DESH. However, this combination could not improve the CSF segmentation accuracy. Another approach is needed to overcome this challenge.

Background

The neuroradiological characteristics of idiopathic normal pressure hydrocephalus (iNPH) are ventricular and Sylvian fissure dilation and a narrow cerebrospinal fluid (CSF) space in the superior convexity [1]. These findings are referred to as disproportionately enlarged subarachnoid space hydrocephalus (DESH) [2]. These severe morphological changes hinder the statistical analyses of brain images because brain segmentation and spatial normalisation errors sometimes occur when the ventricles and the Sylvian fissures are extremely large, and statistical analysis software, such as the Statistical Parametric Mapping 12 (SPM12) software package (http://www.fil.ion.ucl.ac.uk/spm/), is used for brain images. In our previous studies, we found that the dura and/or extradural structures in patients with iNPH were often misidentified as grey matter, white matter and/or CSF, and the anterior and posterior horns of the lateral ventricles were also frequently misidentified as white matter when SPM12 was used [3, 4]. Therefore, we manually stripped the images of the sinuses, dura, and extradural structures and created masks of the ventricles to avoid these segmentation errors. In addition, white matter lesions with hypointensities in T1-weighted images are another characteristic of neuroradiological findings in patients with iNPH and are mostly misidentified as grey matter, which causes mismeasurement of the grey and white matter volumes, especially in periventricular structures [5, 6].

Furthermore, the spatial normalisation accuracy of brain images in patients with iNPH has been inadequate for voxel-based statistical analyses not only in the periventricular and perisylvian areas but also in the superior convexity [4]. When single-photon emission computed tomography (SPECT) perfusion images from patients with iNPH are spatially and statistically normalised using via three-dimensional stereotactic surface projections (3D-SSPs) [7, 8], false hyperperfusion in the superior convexity region is frequently observed [9]. This finding is caused by the misregistration of compressed gyri in this area and is referred to as the convexity apparently hyperperfusion (CAPPAH) sign. In contrast, false hypoperfusion in the perisylvian areas is also often observed [9]. This finding occurs because part of the dilated Sylvian fissures is misidentified as the perisylvian cortex and is referred to as the check mark sign (hypoperfusion area along the Sylvian fissure looks like a check mark). Despite incorrect findings, these signs are paradoxically used for the differential diagnosis of iNPH. Further approaches for more accurate segmentation and spatial normalisation of brain images are needed to perform reliable statistical analyses of brain structures in patients with iNPH.

The Computational Anatomy Toolbox (CAT12) is an advanced tool for morphometric analyses of brain structures [10]. In particular, its segmentation algorithm is unique because it is based on the adaptive maximum a posteriori (AMAP) technique, which does not require a priori information on brain tissue probabilities [10]. Unlike SPM12, tissue probability maps (TPMs) [11] are not constantly used in the process of segmentation but are only used for spatial normalisation, initial skull stripping, and initial segmentation estimation in CAT12. AMAP updates the estimation models for brain tissue classification and addresses the local intensity variation in the original brain images derived from subject-specific and/or MRI unit-specific variation in brain structures. In addition, partial volume estimation (PVE) with a simplified mixed model of a maximum of two tissue types (grey and white matter and grey matter and CSF) calculates the amount (or fraction) of each pure tissue type that is present in each voxel [12]. Furthermore, CAT12 can eliminate white matter hypointensity (WMH) in the periventricular and deep white matter. CAT12 has an original brain atlas and white matter lesion maps [10]. This atlas and these maps are fitted to an individual brain image using diffeomorphic anatomical registration through exponentiated Lie algebra (DARTEL) [13] or geodesic shooting [14] algorithms, and grey matter-like lesions that do not exist in the white matter are detected. Therefore, the algorithm of brain segmentation in CAT12 might have advantages for reliable statistical analyses of brain structures in patients with iNPH.

However, the developers of CAT12 described that the default settings in CAT12 may not be optimal for some cases, such as analysis in children or certain patient groups, and suggested the utility of creating customised TPM and DARTEL or shooting templates for addressing these specific cases (https://neuro-jena.github.io/cat/). The severe morphological changes in the brains of patients with iNPH may also be attributed to atypical brain morphology in these patients. Therefore, we hypothesised that customised TPMs and DARTEL templates can produce more accurate brain image segmentation and spatial normalisation of brain structures in patients with DESH than the default setting of CAT12. Therefore, the aims of this study were as follows: (1) to develop customised TPMs and DARTEL templates (DESH-TPM and DESH-Template) using the brain images of patients with DESH and (2) to elucidate whether the DESH-TPM and DESH-Template can improve the accuracy of brain image segmentation and spatial normalisation. For these purposes, we also segmented and spatially normalised the brain images of healthy elderly controls using CAT12 and referred to the results as standards for the accuracy of these image processes.

Methods

Participants

Two hundred ninety-eight (140 women/158 men) patients with iNPH who were admitted to Tohoku University Hospital were enrolled in this study. The patients were diagnosed by board-certified neurologists on the basis of the diagnostic criteria established according to the Japanese Clinical Guidelines for iNPH [15,16,17]. The inclusion criteria for iNPH patients in this study were as follows: (1) > 60 years of age; (2) gait disturbance, dementia, and/or urinary incontinence measured by the idiopathic normal-pressure hydrocephalus grading scale (iNPHGS) [18]; (3) both ventricular dilatation (Evans’ index > 0.3) and superior convexity and medial subarachnoid space tightness on coronal MRI (DESH); (4) CSF pressure < 200 mm H₂O with normal CSF cell counts and protein levels; (5) the absence of other diseases that may account for such symptoms; and (6) lack of a previous history of illness that might cause ventricular dilatation.

In addition, 25 healthy controls (HCs) were recruited from Tohoku University Hospital (7 women/7 men) or South Miyagi Medical Centre (6 women/5 men). The inclusion criteria for the HCs were as follows: (1) > 55 years of age; (2) lack of a previous history of illness that may cause motor, cognitive, and/or urinary dysfunction; and (3) lack of abnormal findings on neurological examination or brain MRI.

Clinical assessments

In the present study, clinical measures were assessed prior to performing both CSF removal and shunt placement in iNPH patients. In addition to the iNPHGS, we administered the Timed Up and Go test (TUG) [19] for the assessment of gait function and the Mini-Mental State Examination (MMSE) [20] for general cognitive function.

MRI procedure

For all patients with iNPH, cranial MRI was performed using a GE Signa 1.5-Tesla MRI scanner (General Electric Company, Milwaukee, WI, USA) between January 2008 and June 2014; a TOSHIBA Vantage Titan 3-Tesla MRI scanner (Toshiba Medical Systems Corporation, Otawara, Japan) between July 2014 and June 2015; a SIEMENS MAGNETOM Trio, A Tim System 3-Tesla MRI scanner (Siemens Medical Solutions, Erlangen, Germany) between August 2015 and March 2020; or a SIEMENS MAGNETOM Vida 3-Tesla MRI scanner (Siemens Medical Solutions, Erlangen, Germany) between April 2020 and December 2023. For all healthy controls, cranial MRI was also performed using a GE Signa 1.5-Tesla MRI scanner at Tohoku University Hospital or a GE Signa 3-Tesla MRI scanner (General Electric Company, Milwaukee, WI, USA) at South Miyagi Medical Centre.

High-resolution structural image data were acquired via three-dimensional spoiled gradient echo (3D-SPGR) or three-dimensional magnetisation-prepared rapid gradient echo (3D-MPRAGE) imaging sequences. The z-axis of each image in native space was manually oriented parallel to the anterior and posterior commissure (AC-PC) line, and the coordinate of the anterior commissure was also manually set as the origin via SPM12. The imaging parameters used for the acquisition of each MRI scanner are shown in Table S1.

Image data selection and processing

Figure 1 shows the flow chart of patient image data selection and processing. First, the structural images of 260 patients (obtained between January 2008 and March 2020) were selected as the candidate images for developing the DESH-TPM and DESH-Template, and the remaining images of 38 patients (obtained between April 2020 and December 2023) were used to evaluate the validity of brain image segmentation and spatial normalisation when using CAT12 with the DESH-TPM and DESH-Template. Second, ninety-one of the 260 images were excluded for developing the DESH-TPM and DESH-Template because these images included brain infarction and/or haemorrhage, motion and/or metal artefacts, cavum Vergae, focal CSF pooling in the superior convexity, image deficits, brain tumours, and/or subarachnoid cysts. Images with WMHs were included for developing the DESH-TPM and DESH-Template.

Fig. 1

Flow chart of the process of creating the DESH-TPM and DESH-Template CSF: cerebrospinal fluid; DESH: disproportionately enlarged subarachnoid space hydrocephalus; and TPM: tissue probability map

In this study, cranial image segmentation and spatial normalisation were performed only using CAT12. The remaining 169 images were first segmented into grey and white matter, CSF, bone, soft tissues, and other structures using the standard TPM in SPM12 and the DARTEL templates in CAT12. The grey and white matter and CSF images of eighty of 169 patients were correctly segmented, but the images of the remaining 89 patients were incorrectly segmented. The images in which the bone and/or soft tissues were not correctly segmented were not excluded from subsequent processing. Second, we applied the DARTEL algorithm template in SPM12 and created the prototype DESH-Template using well-segmented grey and white matter and CSF images. Third, the images of these 5 structures (grey and white matter, CSF, bone, and soft tissues) were aligned on the basis of the prototype DESH-Template. The mean images of these 5 structures were also used to create the prototype DESH-TPM. Fourth, we created a mask that was fitted to the lateral ventricles and Sylvian fissures in the mean image of CSF, and the mask was expanded into 8 voxels larger from the edge via MRIcron [21]. The mean images of grey and white matter and CSF were smoothed via an isotropic Gaussian filter (8 mm, full width at half maximum) only within the region of the expanded mask to cope with both the morphological diversity within and around the ventricular and Sylvian structures and the accurate segmentation of the superior convexity in patients with iNPH. A TPM used for SPM12 or CAT12 had to be created so that the sum of their probability values for each structure was 1 for each voxel. Therefore, 5 smoothed images and one image, which was calculated by subtracting the sum of the probability values of 5 structures in each voxel from 1, were used for the prototype DESH-TPM.

The eighty-nine patient images, which failed to be correctly segmented when we used the standard TPM in SPM12 and the DARTEL templates in CAT12, were newly segmented via the prototypes DESH-TPM and DESH-Template. Thirty-four of the 89 patient grey and white matter and CSF images were correctly segmented, but the other 55 images were incorrectly segmented. Finally, we created the final DESH-Template and DESH-TPM using images of 5 structures (grey and white matter, CSF, bone, and soft tissues) derived from 114 patients in total. The process was the same as that used to create the prototype DESH-Template and DESH-TPM. The final DESH-TPM is shown in Fig. 2. Unexpectedly, compared with the standard TPM in SPM12, a smaller pontine cistern, thicker skull base, and distorted structures adjacent to the foramen occipital magnum were observed in the DESH-TPM. Consequently, the final DESH-Templates and DESH-TPM were created via 77 (67.5%) images obtained via the 1.5-Tesla GE Signa MRI scanner, 17 (14.9%) images via the 3-Tesla TOSHIBA Vantage Titan MRI scanner, and 20 images (17.5%) via the 3-Tesla SIEMENS MAGNETOM Trio, A Tim System MRI scanner. The demographics of the patients who developed the final DESH-Template and DESH-TPM are shown in Table 1.

Fig. 2

DESH-TPM and SPM12-TPM. a DESH-TPM. b SPM12-TPM. Red area: grey matter; yellow area: white matter; blue area: CSF; white area: bone; and yellow–green area: soft tissues. DESH disproportionately enlarged subarachnoid space hydrocephalus, SPM12 statistical parametric mapping 12, TPM tissue probability map

Table 1 Demographics of participants

Validation study for the DESH-TPM and DESH-Template methods

The demographics of the participants in the validation study are shown in Table 1. All the structural images of 38 patients (obtained between April 2020 and December 2023) were included in the validation study despite apparent brain infarction, focal CSF pooling in the superior convexity, and/or WMH. In these images, haemorrhage, artefacts, cavum vergae, image deficits, brain tumours, and/or subarachnoid cysts were not observed. The patient images were segmented and normalised under two conditions: (1) applying the DESH-TPM and DESH-Template (customised condition); or (2) applying the standard TPM in SPM12 and the DARTEL templates in CAT12 (standard condition). Moreover, the structural images of 25 HCs were also segmented and normalised using the standard TPM and DARTEL templates in CAT12 (reference condition).

Two board-certificated neurologists who specialise in the interpretation of brain images independently examined the error types during segmentation. If the evaluation results of the error types in each patient segmented images differed between the two raters, all the discrepancies were discussed until an agreement was reached. In addition, we calculated the concordance rates between each subject’s normalised grey matter image and the mean grey matter image and between each subject’s normalised white matter image and the mean white matter image in each condition via the three-dimensional-structural similarity (3D-SSIM) algorithm for greyscale images in MATLAB Version 24.1.0.2537033 (R2024a). These concordance rates were used as indices of intrasubject accuracy of normalisation. Moreover, we investigated how correctly the normalised images of grey and white matter were fitted to the DESH-Templates or DARTEL templates in CAT12 under each condition. First, we trimmed the areas whose probability values were less than 0.5 from the normalised grey and white matter images and created binary maps, which indicate the areas with high probabilities of grey and white matter in each subject. These binary images of each structure were overlapped in each condition via MRIcron. On the basis of the overlapped images, we identified the regions that had higher or lower overlap rates in these normalised images. We also calculated the mean overlap rates within and outside the voxels whose probability values were above 0.5 in the highest-resolution DESH-Template of each structure (Template_6.nii in Additional file 2.zip) or standard DARTEL template in CAT12 (Template_6_Dartel.nii) under each condition. We used these variables as the indices of accuracy (within the voxels) and inaccuracy (outside the voxels) of normalisation between the subject normalised images and the templates we applied under each condition.

Additional analyses of CSF image deficits

We frequently observed CSF image deficits in the validation study of the DESH-TPM and DESH-Template methods (please see the Results section). We hypothesised that CSF image deficits may be associated with the inhomogeneous intensities of CSF images. Therefore, we first selected the patient structural images obtained by the SIEMENS MAGNETOM Vida 3-Tesla (highest frequency of Error 3) and GE Signa 1.5-Tesla (lowest frequency of Error 3) MRI scanners. Second, we manually measured the means of CSF intensities via global and local intensity-corrected structural images created during the processing of CAT12 and MRIcron to create each CSF region of interest (ROI) as follows (Figure S2): (1) superior convexity; (2) lateral ventricles; (3) Sylvian fissures; and (4) infratentorial cisterns [22]. We then normalised the mean intensity of each region via CSF intensity estimation derived from the AMAP algorithm during the segmentation of each patient’s brain image. Finally, we evaluated the differences in these normalised mean intensities among the CSF regions at each MRI scanner.

Statistical analyses

The chi-square test and post hoc chi-square test with Bonferroni correction were used to compare the frequencies of each error type during segmentation among the images obtained by the three MRI scanners in the process of creating the DESH-Template and DESH-TPM and among the three conditions in the validation study. In addition, the concordance rates in each condition were compared among the three conditions via one-way analysis of variance (ANOVA), followed by Tukey’s post hoc analysis for comparisons between all pairs of each condition. ANOVA, followed by Tukey’s post hoc analysis, was also used for comparisons of CSF intensity measurements among the four CSF regions and between all pairs of CSF regions at each MRI scanner.

Statistical analyses were performed via IBM SPSS statistics software (Version 25.00; IBM SPSS, Inc., Armonk, NY, USA), and statistical significance was defined as a p value < 0.05.

Results

Error types and their frequencies during segmentation

First, we demonstrated the error types and their frequencies during segmentation when the DESH-TPM and DESH-Template were created. The error types detected during segmentation were as follows. Error 1: misidentifying the dura and/or extradural structures as grey and/or white matter and/or CSF; Error 2: misidentifying white matter lesions (represented by white matter hypointensity) as grey matter; Error 3: CSF-image deficits; Error 4: misidentifying grey matter as the CSF (Figure S1-a); and Error 5: collapsed brain structures (Figure S1-b). The frequency of each error type is shown in Table S2. The frequency of Error 3 significantly differed among the MRI scanners we used. Error 3 was more frequently observed in the segmentation of images obtained by the SIEMENS MAGNETOM Vida 3-Tesla MRI scanner than in those obtained by the other MRI scanners. Error 5 occurred only in the process of creating the prototype DESH-template and DESH-TPM (when the standard TPM in SPM12 and the DARTEL templates in CAT were used).

Table 2; Fig. 3 present the error types and their frequencies in each condition of the validation study. The frequencies of Error 1 under customised, standard, and reference conditions were 10.5%, 44.7%, and 13.6%, respectively. The frequency of Error 1 significantly differed among the three conditions (p = 0.001). The chi-square test with Bonferroni correction (p < 0.05/3) revealed that Error 1 was significantly more common in the standard condition than in the customised (p = 0.001) and reference (p = 0.006) conditions. In contrast, the frequency of Error 1 did not significantly differ between the customised and reference conditions (p = 0.856). Under standard condition, the sagittal sinus, dura, and extradural structures in the superior convexity region were frequently misidentified as grey and/or white matter and/or CSF (Fig. 3a). Moreover, the posterior cranium (sagittal sinus, dura, and extradural structures) was misidentified as grey and/or white matter and/or CSF under all conditions (Fig. 3b).

Table 2 The error types during segmentation and their frequencies in the validation study of DESH-TPM and DESH-Template

Fig. 3

Error types of brain image segmentation in each condition. a demonstrates the misidentification of dural and extradural structures as the grey or white matter or CSF in the superior convexity (yellow arrowhead) and the CSF image deficits in the infratentorial region (white arrowhead). The segmented CSF images are coloured light blue. b shows the misidentification of dura as the grey or white matter in the posterior cranium (red arrowhead). This type of error was observed in all conditions. c demonstrates the errors of WMH cancellation. The errors in the periventricular white matter were frequently observed in the standard condition (yellow green arrowhead). Errors in the frontal subcortical white matter were identified in both the customised and standard conditions (pink arrowhead). d shows the CSF image deficits in the superior convexity and Sylvian fissures (light blue arrowhead). The segmented CSF images are coloured light blue. The CSF image deficits in the superior convexity were observed in all conditions. CSF image deficits in the Sylvian fissures were not identified; only in the reference condition

The frequencies of Error 2 under customised, standard, and reference conditions were 18.4%, 44.7%, and 0%, respectively. The frequency of Error 2 was significantly different among the three conditions (p < 0.001). The chi-square test with Bonferroni correction (p < 0.05/3) revealed that Error 2 was significantly more common in the standard condition than in the customised (p = 0.001) and reference (p < 0.001) conditions. In addition, Error 2 tended to be more common in the customised condition than in the reference condition (p = 0.023). The misidentification of periventricular WMH as grey matter was frequently observed under standard conditions, whereas both customised and standard conditions misidentified lesions with hypointensity in the subcortical white matter of the frontal lobes as grey matter (Fig. 3c).

In contrast, the frequencies of Error 3 under customised, standard, and reference conditions were 97.4%, 84.2%, and 28%, respectively. The frequency of Error 3 was significantly different among the three conditions (p < 0.001). The chi-square test with Bonferroni correction (p < 0.05/3) revealed that the frequency of Error 3 in the reference condition was significantly lower than those in the customised (p < 0.001) and standard (p < 0.001) conditions. Although Error 3 tended to be more frequent in the customised condition than in the standard condition, this difference was not significant (p = 0.047). CSF image deficits in superior convexity were observed in all conditions (customised condition: 71.1%; standard condition: 84.2%; and reference condition: 28%). CSF image deficits in the anterior parts of Sylvian fissures were frequently observed under customised and standard conditions (Fig. 3d). Moreover, deficits in CSF images of the prepontine cistern and cisterna magna were more frequently observed under customised conditions (92.1%) (Fig. 3a). These errors in the infratentorial CSF spaces were uncommon under standard conditions (10.5%) and were not observed under reference conditions. All the regions with CSF-image deficits were misidentified as bone.

Errors 4 and 5 were not observed under any conditions.

Analyses of CSF intensities

Table S3 shows the results of the CSF intensity measurements. ANOVA revealed significant differences in the normalised mean intensities among the four CSF regions at each MRI scanner [SIEMENS MAGNETOM Vida 3-Tesla: F (3, 148) = 719.56, p < 0.001; GE signa 1.5T: F (3, 304) = 73.981, p < 0.001]. In the SIEMENS MAGNETOM Vida 3-Tesla MRI scanner, Tukey’s post hoc analysis revealed that the normalised mean intensities in the superior convexity region were the lowest among the four regions (p values of all comparisons were less than 0.001), and those in the infratentorial cisterns were the highest (p values of all comparisons were less than 0.001). In particular, the mean of the normalised mean intensities in the superior convexity was more than 2 standard deviations (SDs) below from the mean intensity of the whole CSF images estimated by the AMAP algorithm. The normalised mean intensities did not exceed + 2 SDs in any region above the mean intensity. All original images obtained by a SIEMENS MAGNETOM Vida 3-Tesla MRI scanner had central brightening artefacts (Figure S3) [23,24,25].

In contrast, with a GE Signa 1.5-Tesla MRI scanner, Tukey’s post hoc analysis revealed that the normalised mean intensities in the superior convexity and Sylvian fissures were significantly greater than those in the other regions (p < 0.001). The normalised mean intensities in the infratentorial cisterns were significantly greater (p < 0.001) than those in the lateral ventricles. There was no significant difference in the normalised mean intensities between the superior convexity and Sylvian fissures (p = 0.976). The mean of the normalised mean intensities was not below − 2 SDs or above + 2 SDs from the mean intensity of the whole CSF images estimated by the AMAP algorithm in any region.

Accuracy of normalisation

The concordance rates and the mean overlap rates of the normalised grey and white matter images are shown in Table 3. ANOVA revealed significant differences in the concordance rates of the normalised grey and white matter images among the three conditions [grey matter: F (2, 98) = 161.85, p < 0.001; white matter: F (2, 98) = 116.60, p < 0.001]. Tukey’s post hoc analysis demonstrated that the customised condition had the highest concordance rate of the normalised grey and white matter images among the three conditions (p values of all comparisons were less than 0.001). Compared with the reference condition, the standard condition had a significantly lower concordance rate of the normalised grey matter images (p = 0.001), whereas the concordance rate of the white matter images did not significantly differ between the standard and reference conditions (p = 0.775).

Table 3 Results of concordance rates under each condition

The mean overlap rates of the normalised grey matter images within and outside the voxels whose probability values were above 0.5 in the grey matter template were 84.5% (accuracy index) and 8.1% (inaccuracy index) in the customised condition, 75.2% and 15.3% in the standard condition, and 79.5% and 13.9% in the reference condition, respectively. The mean overlap rates of the normalised white matter images within and outside the voxels whose probability values were above 0.5 in the white matter template were 91.0% (accuracy index) and 4.8% (inaccuracy index) in the customised condition, 83.4% and 7.6% in the standard condition, and 87.2% and 8.8% in the reference condition, respectively.

Figures 4 and S4 show the coloured overlapping maps of normalised grey and white matter images under the three conditions. Overall, the overlapping maps of the normalised grey and white matter images in the customised condition were the most highly fitted to each structural template among the three conditions (Figure S4). In particular, for superior convexity, the overlapping map of normalised grey matter images in the standard condition was most poorly fitted to the grey matter template among the three conditions (Fig. 4a). In contrast, the overlapping map of the normalised white matter images in the customised condition was the most highly fitted to the white matter template among the three conditions (Fig. 4b).

Fig. 4

Overlapped maps of normalised grey and white matter images in three conditions. a shows the overlapped normalised grey matter image. b shows the overlapped normalised white matter image. Coloured maps demonstrate the overlap rates of each structural images outside the voxels of which probability values were above 0.5 in each structural template, which indicates the magnitude of spatial normalisation inaccuracy

Discussion

In the present study, we developed the DESH-TPM and DESH-Template and validated their utility for brain structure segmentation and spatial normalisation when using CAT12 in patients with DESH. The major findings of the study were as follows: (1) applying the DESH-TPM and DESH-Template to brain images in patients with DESH can improve the segmentation and spatial normalisation accuracy of grey and white matter compared with the standard TPM in SPM12 and the Dartel templates in CAT12; and (2) the CSF segmentation accuracy of brain images in patients with DESH was not improved when the DESH-TPM and DESH-Template were used.

Grey and white matter segmentation

The results of our study revealed that the DESH-TPM and DESH-Template methods could improve the segmentation accuracy of grey and white matter, especially in terms of superior convexity. The main reason for this improvement was the reduced frequency of misidentifying the dura, sinus, and extradural structures as grey and/or white matter. The DESH-TPM and DESH-Template could also improve the accuracy of WMH cancellation in the periventricular white matter. As we hypothesised, these findings suggested that the DESH-TPM could provide a more accurate probability of each brain structure than the standard TPM in SPM12 when applied to patients with DESH. Moreover, the brain regions and WMH atlases in CAT12 could be more correctly fitted to at least the superior convexity and periventricular white matter regions of each patient’s brain image by using DESH-Template because the adaptation of these maps to each patient’s brain image was performed on the basis of the applied DARTEL templates in each condition [10].

Moreover, the posterior cranium was not accurately segmented under any condition of CAT12 processing. The posterior cortexes were mostly adjacent to the dura and sinuses, and the CSF space around the posterior cortexes was very narrow in both patients with DESH and healthy controls. In addition, the intensities of grey and white matter are relatively close to those of the dura and sinuses. Therefore, this segmentation error seems to occur fundamentally because of the CAT12 algorithm and cannot be avoided despite the application of the DESH-TPM and DESH-Template methods to the brain images of patients with DESH. Moreover, lesions with hypointensities in the subcortical white matter of the frontal lobes were misidentified as grey matter under customised and standard conditions. This error was attributed to the WMH atlas in CAT12 because the WMH probabilities in the subcortical white matter are originally lower than those in the periventricular white matter. Additional approaches, such as a combination of structural and fluid attenuated inversion recovery images, are needed to prevent these segmentation errors [26, 27].

CSF segmentation

Unfortunately, the accuracy of CSF segmentation for superior convexity and Sylvian fissures did not improve when DESH-TPM and DESH-Template were used. We observed CSF image deficits in superior convexity in any condition, and those in the Sylvian fissures were identified only in customised and standard conditions (brain images of patients with DESH). The regions with CSF image deficits were misidentified as bone in all patients. One possible reason for these CSF image deficits is the strong adaptation of the original brain atlas in CAT12 to each patient’s brain image. CAT12 adopts the DARTEL or shooting algorithm for adaptation, and this process is associated with brain tissue classification and local intensity variation. CSF image deficits in the superior convexity frequently occurred in the focally dilated sulci, and those in the Sylvian fissures were mostly observed in the anterior parts, which were more dilated than those in the posterior parts. The adaptation of the original brain atlas based on DARTEL or shooting algorithms seems to contribute to accurate grey and white matter segmentation in superior convexity and perisylvian structures, whereas it may easily erode the outlines of intracranial structures in these areas. Moreover, CSF image deficits in the prepontine cistern and cisterna magna were observed mostly under customised condition. This error was attributed to the smaller pontine cistern, thicker clivus, and distorted images of the foramen occipital magnum and the surrounding structures in the DESH-TPM. The DESH-TPM was created primarily using images obtained by a GE Signa 1.5-Tesla MRI scanner. Unfortunately, almost all of the images included brightening and distortion artefacts around the foramen occipital magnum, which deteriorated the quality of DESH-TPM.

However, these CSF image deficits were uncommon in structural images obtained via a GE Signa 1.5-Tesla or a TOSHIBA Vantage Titan 3-Tesla MRI scanner despite the application of the prototype DESH-TPM to the brain images of patients with DESH. We found that the CSF image intensities obtained by a SIEMENS MAGNETOM Vida 3-Tesla MRI scanner with superior convexity were extremely low, and those in the infratentorial cisterns were relatively high compared with those in the other CSF regions, despite the global and local correction of the intensities of these images by CAT12. Several previous MRI studies reported central brightening artefacts, which manifested as high signal intensity in the centre of images [23,24,25]. Bernstein et al. reported that these artefacts can occur when the radiofrequency wavelengths approximate the sizes of the observation objects [25]. These shorter wavelengths can spatially modulate the image intensity, and central brightening artefacts were more frequently identified when 3-Tesla than when 1.5-Tesla MRI scanners were used because the radiofrequency wavelengths at 3-Tesla were approximately two times shorter than those at 1.5-Tesla. Keihaninejad et al. reported lower CSF image intensities in the cisterns than in the ventricles at 3-Tesla and suggested that this discrepancy led to the underestimation of CSF spaces in the cisterns [28]. They concluded that their CSF space measurement was likely to underestimate the real intensity discrepancy between the ventricles and peripheral cisterns because they included the cisterns located on a central part of the brain, such as the prepontine cistern, to measure the CSF intensities. The CSF-image intensities in these cisterns apparently increased because of central brightening artefacts. These findings in Keihaninejad et al.’s study are consistent with the features of brain images at 3-Tesla in our study. Although CAT12 processing has an algorithm to remove image-intensity inhomogeneities, the CSF-image inhomogeneities at 3-Tesla could not be sufficiently removed to perform accurate CSF-image segmentation. We speculated that not only poor probability information of the infratentorial space in the DESH-TPM but also lower CSF image intensities in the superior convexity and higher CSF image intensities in the infratentorial cisterns facilitate the TPM-dependent trimming of intracranial structures, which erodes the outlines of the cisterns.

Normalisation

DESH-Template improved the spatial normalisation accuracy of grey and white matter images in patients with DESH. This improvement was primarily due to the higher concordance rates of the normalised images at the superior convexity. This finding indicates greater localisation homogeneity of brain structures in superior convexity in patients with DESH. Although the normalised grey matter in the medial superior convexity was misaligned to more medial areas in the standard condition, this misalignment was not identified in the reference condition. A previous voxel-based morphometry study revealed that the grey matter density in the superior convexity region in patients with iNPH was greater than that in healthy controls, and this finding was especially noticeable in the medial part of the superior convexity region [29]. However, our results suggest that the findings of the previous study did not reflect higher grey matter density but were derived from the misalignment of normalised grey matter in the medial part of superior convexity. ROI-based analysis may help elucidate whether patients with DESH truly lack atrophy in the superior convexity. Therefore, we plan to develop a brain-region atlas for DESH-Template.

Moreover, we could not detect apparent differences in local concordance of normalisation except for the superior convexity among the three conditions. This finding suggests that the standard templates in CAT12 and the DESH-Template performed similarly well in all brain structures except for the superior convexity. However, we detected higher rates of WMH misidentification as grey matter, grey matter misidentification as the CSF, and collapsed brain structures during segmentation when the standard TPM and DARTEL templates in CAT12 were used. In the segmentation processing of CAT12, the selected DARTEL or shooting templates are used to fit the brain atlas to subject brain images [10]. Therefore, the selected templates affect the accuracy of brain segmentation. Our results suggest that combining the DESH-TPM and DESH-Template has advantages for the accurate segmentation and spatial normalisation of all brain images except for those of the CSF in patients with DESH.

Limitations

Our study has several limitations. First, standard maps for ROIs, such as the Anatomical Automatic Labelling Atlas (AAL) [30], cannot be used for ROI-based analyses if the DESH-Templates are used for the spatial normalisation of brain images. Therefore, the original brain-region map, which is fitted to the DESH-Template, needs to be developed for performing ROI-based analyses. We are now developing a brain-region atlas for DESH-Template according to the brain regions in the AAL and validating its reliability and utility.

Second, accurate measurements of intracranial volumes cannot be provided if the DESH-TPM is applied to segment cranial images because our study suggests that CSF image deficits may frequently occur when we apply the DESH-TPM to cranial images obtained by a high-field MRI scanner. Furthermore, CSF image deficits in the cisterns observed when default settings were used have also been reported [31]. Therefore, we are currently also developing an algorithm to complement CSF image deficits after the processing of brain image segmentations derived from CAT12 with the DESH-TPM and DESH-Template.

Finally, we could not use many images obtained by the SIEMENS MAGNETOM Trio, A Tim System 3-Tesla MRI scanner, for creating the DESH-TPM and DESH-Template because of CSF image deficits. Therefore, the DESH-TPM was created primarily via images obtained via the GE Signa 1.5-Tesla MRI scanner. This may be a disadvantage for not only CSF image segmentation in the prepontine cistern and cisterna magna but also accurate segmentation of grey and white matter. Therefore, we will update the DESH-TPM and DESH-Template after developing the algorithm to complement CSF image deficits.

Conclusions

Our study demonstrated that applying the DESH-TPM and DESH-Template to CAT12 could improve the accuracy of brain image segmentation and the spatial normalisation accuracy of grey and white matter in patients with DESH. However, our approach could not improve the CSF segmentation accuracy of intradural images in these patients. CSF-intensity inhomogeneity is identified when high-field MRI scanners are used, and the low versatility of DESH-TPM for infratentorial structures is a challenge associated with our approach for accurate CSF segmentation. The goals of our future studies are overcoming these challenges and developing an original brain-region atlas for DESH-Templates, which can lead to reliable voxel-based and ROI-based statistical analyses for patients with DESH.