Many degenerative and developmental neurologic diseases are accompanied by subtle changes to brain morphology. Atlas-based computer-assisted morphometry is widely adopted in human neuroimaging, both in research and clinical practice. Atlas-based techniques facilitate rapid analysis of large samples and allow the detection of subtle morphological changes in individuals that might otherwise go un-noticed.12 Morphometric techniques in canine brain research generally use linear3,4 and area4–7 measurements or volumetry performed by manual segmentation of structures.6,8–13 Reports of the use of atlas-based computer-assisted morphometry are uncommon in the veterinary literature and include 4 reports in nonveterinary journals; 2 describe the normal anatomy of the canine brain,14,15 1 describes the use of morphometry to assess gray and white matter atrophy in a canine model for aging,16 and the remaining report17 describes the use of functional MRI to evaluate the olfactory system of clinically normal dogs. The application of atlas-based computer-assisted morphometry in veterinary medicine is currently limited because population-based MRI brain atlases of healthy dogs with various-shaped calvaria are not readily available, translational adaptation of human-based modeling techniques in dogs is difficult, and veterinarians have limited access to the high-performance computers required and are unfamiliar with the software available for this technique.
An MRI brain atlas is defined as a template brain image and its segmentations or labels.18 The MRI images that make up a brain atlas should be acquired at a high-field strength of ≥ 1.5 T with 3-D, high-resolution T1-weighted spoiled gradient or magnetization-prepared rapid-acquisition gradient recalled echo sequences, preferably with ≤ 1 mm3 of isotropic voxels.1 Templates are created by the formation of a composite of high-resolution volumetric MRI sequences from multiple individuals. Prior to the creation of a template, the skull and all soft tissues other than the brain must be removed from the images (ie, the brain must be extracted) to provide a volumetric brain image. Atlas-based segmentation involves coregistration of an MRI brain image from an individual to a template brain image, followed by application of template segmentation. Then, inverse registration is performed to obtain the measurements of interest. Templates may be specific, derived from a narrow sample of similar individuals, or may be general, derived from a broader sample of individuals with varying age, sex, and ethnicity or breed. Specific templates provide high-fidelity registration for similar subjects but have limited applicability. Generalized templates allow registration of a broader range of subjects; however, accuracy of registration may be reduced from that of specific templates.
The objective of the study reported here was to develop a method for constructing an MRI atlas of the canine brain that is representative of the diversity of the general dog population and evaluate 3 methods of ABS. Specific aims were to construct several canine MRI brain atlases that were representative of the brain shapes of various breeds, develop a technique for the automatic extraction of the brain from MRI images to improve the speed of ABS, and validate the repeatability of the method identified as the most accurate ABS technique. Our hypotheses were that ABS used with a template of a brain that has a shape similar to that of the individual being evaluated would be more accurate than ABS used with a unified brain template that is comprised of all brain shapes, that ABS on images from which the brain was automatically extracted would be as accurate as or more accurate than ABS on images from which the brain was manually extracted, and that ABS would be at least as repeatable as manual segmentation, the gold standard.
Materials and Methods
Animals
Hospital records from June 2011 through May 2013 were reviewed to identify MRI brain scans for dogs that did not have epilepsy or evidence of structural brain disease as determined by at least 1 veterinary radiologist (MEM) and the attending veterinary neurologist. Necropsy and histologic examination of brain tissue specimens were not required for dogs to be included in the study. During that period, 44 dogs (template group) met the criteria for study inclusion, and the MRI scans from those dogs were used to construct the MRI templates. Hospital records from June 2013 through June 2014 were reviewed to identify dogs similar to those in the template group, and the MRI scans from those dogs (n = 18; test group) were used to validate 3 ABS methods. All dogs underwent complete physical and neurologic examinations. Bloodwork such as a CBC and serum biochemical analysis and CSF analysis were performed intermittently at the discretion of the attending clinician and were not used as selection criteria for the study. All dogs were examined as part of another research project12 that was approved by the University of Melbourne Animal Ethics Committee (n = 20) or were examined as part of a diagnostic workup with the owners' consent and were exempt from oversight by the animal ethics committee.
MRI
All dogs were imaged with a 1.5-T MRI scannera by use of either a knee coil or 8-channel brain coil.a Multiple sequences were acquired in accordance with the institution's standard brain imaging protocol and included the sagittal plane T2-weighted fast spin echo, transverse plane T2-weighted fluid-attenuated inversion recovery, T2* gradient echo, diffusion-weighted imaging with calculation of the apparent diffusion coefficient map, and dorsal plane T1-weighted 3-D fast spoiled gradient recalled (3-D ultrafast gradient echo) sequence acquired before and after administration of gadodiamideb (0.1 mmol/kg, IV). For this study, the only sequence of interest was the T1-weighted 3-D fast spoiled gradient recalled sequence with the following settings: slice thickness, 1 mm; repetition time, 13 milliseconds; echo time, 5.7 milliseconds; inversion time, 600 milliseconds; number of excitations, 2; and flip angle, 15°. The acquisition time for this sequence ranged from 8 to 12 minutes and was dependent on the required number of phase-encoding steps.
Classification of dogs by brain shape
The ability to successfully build a template depends on the similarity in the overall shape of the constituent brains. Therefore, the dogs of the template group were classified as having a brachycephalic, mesaticephalic, or dolichocephalic brain shape on the basis of subjective evaluation of the MRI scans by the primary investigator (MEM), regardless of the recorded breed for each dog. Brains that were round and in which the bulk of the olfactory bulb was rotated ventral to the frontal lobes were classified as having a brachycephalic shape, whereas those that were less round in which the bulk of the olfactory bulb was located equal or rostral to the frontal lobes were classified as having a mesaticephalic shape, and those that were elongated ovals with prominent olfactory bulbs that extended rostral to the frontal lobes were classified as having a dolichocephalic shape. Currently, there is no consensus for the classification of dogs on the basis of brain shape. Application of the cephalic index19 or an adapted brain index to the dogs of the template group did not provide clear separation into 3 groups (Online Supplement available at http://avmajournals.avma.org/toc/ajvr/77/4). The objective measurement that correlated best with the subjective classification of brain shape was brain length. Thus, for objectivity during the validation phase of the study, dogs in the test group were classified in accordance with brain length. Dogs with a brachycephalic shape had a brain length < 68 mm, whereas those with a mesaticephalic shape had a brain length between 72 and 87 mm, and those with a dolichocephalic shape had a brain length > 88 mm. In the test group, 6 dogs were classified in each brain shape category.
Construction of brain atlases
Each dog's DICOM (Digital Imaging and Communications in Medicine) files were imported into a high-performance computerc and converted to the neuroimaging informatics technology initiative file format for all manipulations. Most manipulations were run from the command line with the computer programming language of bash scripts. Automatic brain extraction was attempted by use of the brain extraction tool in open-source image analysis software20,d; however, that tool is tailored for extraction of human brains and did not provide adequate extraction of canine brains. Thus, each brain was manually extracted from the skull in each image by tracing a brain mask by hand, slice by slice, with open-source segmentation software.21,e The mask was binarized and inflated, then subtracted from the whole head image to produce the extracted brain image. All manual extractions were performed by the same investigator (MEM). Unbiased templates were generated from the T1-weighted neuroimaging informatics technology initiative brain images with open-source multidimensional image registration and segmentation softwaref with buildtemplateparallel script as described.22–24 That script performs an affine registration of the input images into 1 brain and intensity and averages the result to produce a new, fuzzy-looking brain image. Each of the original images was then nonlinearly registered by the use of an SyN algorithm to the averaged brain image to produce a new shape-based average brain. The nonlinear registration was repeated 4 times. Each iteration was used to update the average brain image until an optimal study-specific template was obtained. The script was run on a high-performance 64-core computerc and was completed within 4 hours. Preceding all image registrations, N4 bias field corrections were performed with default settings (cross-correlation similarity metric, 60 × 90 × 40 nonlinear iterations by use of a Greedy-SyN transformation model with a 0.2 step-size).
Templates were constructed for each of the 3 brain shapes, and a combined template was constructed for the mesaticephalic and dolichocephalic shapes, resulting in the creation of 4 templates. Attempts to construct a unified template comprised of all 3 brain shapes were unsuccessful. For each of the 4 templates, structures were manually segmented by the primary investigator (MEM), who traced the structure margins slice by slice with segmentation softwared to produce labels for the brain, left and right hippocampal formations, left and right caudate nuclei, and ventricular system. Those structures were chosen because they have fairly well-defined boundaries on individual conventional MRI images and could be manually segmented with confidence and were therefore useful for evaluation of accuracy. Furthermore, the hippocampal formations and caudate nuclei are structures of clinical interest because they become atrophied in human patients with epilepsy, stroke, and neurodegenerative diseases such as dementia, Alzheimer disease, and Huntington disease1 and might be likewise affected in dogs with epilepsy, cerebral infarction, and canine cognitive dysfunction syndrome.
ABS for dogs of the test group
Atlas-based segmentation is a 2-step process that requires extraction of the brain followed by the application of the desired template to segment out the structure of interest. Three methods of ABS were evaluated. Method A involved manual brain extraction followed by application of 1 of 3 brain shape–specific templates. Method B involved automatic brain extraction followed by application of 1 of 3 brain shape-specific templates. Method C involved manual brain extraction followed by application of the combined template for mesaticephalic and dolichocephalic brain shapes.
The MRI scans for all dogs in the test group underwent both automatic and manual brain extraction. Automatic brain extraction involved registration of the whole head image to a template head image and application of a generous mask based on the template head image, followed by inverse registration to produce a coarsely extracted brain. That technique was rapidly performed and resulted in the removal of most tissues other than the brain from the images. Manual brain extraction was performed in the same manner as that described to produce the templates. Open-source multidimensional image registration and segmentation softwaref was used to perform the registration and inverse registration steps in the same manner as that described to produce the templates.
Evaluation of accuracy and repeatability of ABS
For each of the 18 dogs in the test group, manual segmentation and ABS were performed on the brain, left and right hippocampal formations, and left and right caudate nuclei. The accuracy and repeatability of the results obtained by ABS were compared with those obtained by manual segmentation, which was considered the gold standard. The Dice and Jaccard similarity coefficients, 2 measures of overlap, were calculated as described25 by use of the open-source multidimensional image registration and segmentation software.f Both coefficients compare the extent of spatial overlap between a source volume (S) and a target volume (T) and are expressed as a ratio, with a value of 1 indicating perfect unity (ie, overlap) of the volumes. The Dice coefficient (2 × [|S ∩ T| / |S| – |T|]) is commonly used in human medicine and provided a useful measure for the accuracy of ABS when applied in dogs. The Jaccard coefficient (|S ∩ T| / |S U T|) is more amenable to multivariable analyses and was used in the logistic regression analysis.
Statistical analysis
For both the template and test groups, the respective data distributions for age and body weight were assessed for normality by visual examination of histograms and estimation of the skewness, kurtosis, and Shapiro-Wilk test statistic. The mean age and body weight were determined for both groups and compared by means of a 2-tailed unpaired Student t test for equal or unequal variances as appropriate. The sex distribution was compared between the 2 groups by means of a χ2 test. For all analyses, values of P ≤ 0.05 were considered significant.
Accuracy of the 3 methods of ABS was evaluated by means of a logistic regression model for grouped data with the Jaccard coefficient used as the outcome variable. Predictor variables evaluated in the model included segmentation method (A, B, or C), brain shape (brachycephalic, mesaticephalic, or dolichocephalic), anatomic structure (left or right hippocampal formation, left or right caudate nucleus, or brain), sex, age, and body weight. Univariable analyses were performed initially, and variables with values of P ≤ 0.20 were eligible for inclusion in a multivariable model. Biologically plausible 2-way interaction terms were also evaluated in the multivariable model. Stepwise backward elimination was used to construct the final multivariable model, in which only variables with values of P ≤ 0.05 were retained. The residual deviances were plotted and the Hosmer-Lemeshow goodness-of-fit test was performed to determine how well the model fit the data. For interpretation, logistic regression outputs (reported as log odds) were converted to probabilities, and the differences of proportions (which were equivalent to the marginal mean differences for the Jaccard coefficients) and associated 95% CIs were reported while controlling for all other predictor variables retained in the final model. The Dice coefficients for each method of ABS were calculated and reported to allow for comparisons with ABS methods used in human medicine.
The most accurate ABS method identified during logistic regression was applied to a data set that consisted of a subset of 6 dogs from the test group (2 dogs with each brain shape). Manual segmentation was performed twice on the same subset of 6 dogs, with a minimum of 4 weeks between rounds. The repeatability of each method of segmentation was determined by calculating the Dice coefficient from volumes determined during rounds 1 and 2. The repeatability of the ABS method was compared with manual segmentation.
Graphs were created with a commercially available statistical graphics program.g Logistic regression analyses were performed with a commercially available statistics software program,h and all other analyses were performed with another commercially available statistics software program.i
Results
Template group and construction of brain atlases
Of the 44 dogs in the template group, 10 had a brachycephalic brain shape, 11 had a mesaticephalic brain shape, and 23 had a dolichocephalic brain shape. The dogs with a brachycephalic brain shape included 5 castrated males, 1 sexually intact female, and 1 spayed female and had a mean age of 6.1 years (range, 0.35 to 12.5 years) and body weight of 6.8 kg (range, 1.2 to 13.7 kg). Breeds represented in the brachycephalic brain shape group included Chihuahua, Pug, Shih Tzu, Maltese, Toy Poodle, Cavalier King Charles Spaniel, and mixed breed (n = 4; Pembroke Welsh Corgi crossbred, Pug-Maltese crossbred, Shih Tzu–Maltese crossbred, and Maltese-Poodle crossbred). Of those 10 dogs, 4 had clinical signs of vestibular disease, 1 had a behavioral change, 1 had cervical disc disease, 1 had ataxia, 1 had masticatory myositis, 1 had otitis media, and the diagnosis was undetermined for 1 dog. The dogs with a mesaticephalic brain shape included 4 castrated males and 7 spayed females with a mean age of 9.0 years (range, 3.0 to 16.0 years) and body weight of 23.0 kg (range, 14.9 to 38.6 kg). Breeds represented in the mesaticephalic brain shape group included Cocker Spaniel (n = 2), Boxer (3), mixed breed (3; Kelpie crossbred, Australian Cattle Dog crossbred, and an unspecified crossbred), and Golden Retriever, Shetland Sheepdog, and Whippet (1 each). Of those 11 dogs, 4 had clinical signs of vestibular disease, 2 had a behavioral change, 1 had degenerative myelopathy, 1 had otitis media, 1 had a nonbrain tumor, 1 had vomiting, and 1 had trigeminal neuritis. The dogs with a dolichocephalic brain shape included 5 sexually intact males, 9 castrated males, 3 sexually intact females, and 5 spayed females; the sex for 1 dog was not recorded in the medical record. Those dogs had a mean age of 7.0 years (range, 3.5 to 15.0 years) and body weight of 31.5 kg (30.0 to 38.2 kg). Breeds represented included Greyhound (n = 14), mixed breed (5; 2 Golden Retriever crossbreds, 2 Labrador Retriever crossbreds, and 1 Belgian Sheepdog crossbred), Labrador Retriever (2), Vizsla (1), and Staffordshire Bull Terrier (1). Of those 23 dogs, 14 were clinically normal, 4 had clinical signs of vestibular disease, 1 had a behavioral change, 1 had idiopathic cerebellitis, 1 had hypertension, 1 had immune-mediated polyarthropathy, and 1 had cranial nerve polyneuropathy.
The MRI scans from the 44 dogs in the template group were used to create 4 brain atlas templates, 1 for each of the 3 brain shapes (brachycephalic, mesaticephalic, and dolichocephalic) and a combined template for the mesaticephalic and dolichocephalic brain shapes. Representative images from each of the 4 templates were provided and include segmentations of the hippocampal formation, caudate nuclei, and ventricular system with corresponding 3-D images of the volumes of those structures (Figure 1).
Representative dorsal plane MRI atlas templates of the canine brain for dogs with brachycephalic (rounded brain in which the bulk of the olfactory bulb was rotated ventral to the frontal lobes; A and E), mesaticephalic (brain in which the bulk of the olfactory bulb was located equal or rostral to the frontal lobes; B and F), dolichocephalic (brain with an elongated oval shape and a prominent olfactory bulb that extended rostral to the frontal lobes; C and G), and combined mesaticephalic and dolichocephalic (D and H) brain shapes constructed from the MRI scans of 44 dogs without clinical signs of epilepsy and without MRI evidence of structural brain disease (template group). In panels E through H, segmentation of the ventricular system (blue), left hippocampal formation (yellow), right hippocampal formation (green), left caudate nucleus (red), and right caudate nucleus (pink) has been performed. Panels I through L represent the 3-D renderings of the segmented structures as viewed from the ventral aspect for dogs with brachycephalic (I), mesaticephalic (J), dolichocephalic (K), and combined mesaticephalic and dolichocephalic (L) brain shapes. Notice the high signal-to-noise ratio for each template with excellent visualization of the boundary between the gray and white matter. Also, notice the difference in the shape of the brain among the templates; in particular, the brachycephalic template and associated structures are broader and the hippocampal formation and lateral ventricles have a more upright orientation, compared with the templates for the other brain shapes.
Citation: American Journal of Veterinary Research 77, 4; 10.2460/ajvr.77.4.395
Representative dorsal plane MRI atlas templates of the canine brain for dogs with brachycephalic (rounded brain in which the bulk of the olfactory bulb was rotated ventral to the frontal lobes; A and E), mesaticephalic (brain in which the bulk of the olfactory bulb was located equal or rostral to the frontal lobes; B and F), dolichocephalic (brain with an elongated oval shape and a prominent olfactory bulb that extended rostral to the frontal lobes; C and G), and combined mesaticephalic and dolichocephalic (D and H) brain shapes constructed from the MRI scans of 44 dogs without clinical signs of epilepsy and without MRI evidence of structural brain disease (template group). In panels E through H, segmentation of the ventricular system (blue), left hippocampal formation (yellow), right hippocampal formation (green), left caudate nucleus (red), and right caudate nucleus (pink) has been performed. Panels I through L represent the 3-D renderings of the segmented structures as viewed from the ventral aspect for dogs with brachycephalic (I), mesaticephalic (J), dolichocephalic (K), and combined mesaticephalic and dolichocephalic (L) brain shapes. Notice the high signal-to-noise ratio for each template with excellent visualization of the boundary between the gray and white matter. Also, notice the difference in the shape of the brain among the templates; in particular, the brachycephalic template and associated structures are broader and the hippocampal formation and lateral ventricles have a more upright orientation, compared with the templates for the other brain shapes.
Citation: American Journal of Veterinary Research 77, 4; 10.2460/ajvr.77.4.395
Representative dorsal plane MRI atlas templates of the canine brain for dogs with brachycephalic (rounded brain in which the bulk of the olfactory bulb was rotated ventral to the frontal lobes; A and E), mesaticephalic (brain in which the bulk of the olfactory bulb was located equal or rostral to the frontal lobes; B and F), dolichocephalic (brain with an elongated oval shape and a prominent olfactory bulb that extended rostral to the frontal lobes; C and G), and combined mesaticephalic and dolichocephalic (D and H) brain shapes constructed from the MRI scans of 44 dogs without clinical signs of epilepsy and without MRI evidence of structural brain disease (template group). In panels E through H, segmentation of the ventricular system (blue), left hippocampal formation (yellow), right hippocampal formation (green), left caudate nucleus (red), and right caudate nucleus (pink) has been performed. Panels I through L represent the 3-D renderings of the segmented structures as viewed from the ventral aspect for dogs with brachycephalic (I), mesaticephalic (J), dolichocephalic (K), and combined mesaticephalic and dolichocephalic (L) brain shapes. Notice the high signal-to-noise ratio for each template with excellent visualization of the boundary between the gray and white matter. Also, notice the difference in the shape of the brain among the templates; in particular, the brachycephalic template and associated structures are broader and the hippocampal formation and lateral ventricles have a more upright orientation, compared with the templates for the other brain shapes.
Citation: American Journal of Veterinary Research 77, 4; 10.2460/ajvr.77.4.395
Comparison of the test group with the template group
The test group consisted of 18 dogs that were selected for study enrollment on the basis of the same criteria used for selection of the dogs for the template group. Of the 18 test dogs, 6 were classified in each of the 3 brain shapes. Within each brain shape category, the mean age and body weight for the test dogs did not differ significantly from those for the template dogs (Figure 2). The sex distribution also did not differ significantly between the test and template dogs within the brachycephalic (P = 0.92), mesaticephalic (P = 0.90), and dolichocephalic (P = 0.27) brain shape categories.
Box-and-whisker plots of the data distributions for age (A) and body weight (B) for dogs in the template group of Figure 1 and 18 similar dogs that comprised a test group after the dogs were classified as having a brachycephalic (template group, n = 10; test group, 6), mesaticephalic (template group, 11; test group, 6), or dolichocephalic (template group, 23; test group, 6) brain shape. For each plot, the horizontal line within the box represents the median; the lower and upper edges of the box delimit the 25th and 75th percentiles, respectively; and the whiskers delimit the 5th and 95th percentiles. Within each brain shape, results of 2-tailed Student t tests indicated that neither age nor body weight differed significantly (P > 0.05) between dogs in the template and test groups. See Figure 1 for remainder of key.
Citation: American Journal of Veterinary Research 77, 4; 10.2460/ajvr.77.4.395
Box-and-whisker plots of the data distributions for age (A) and body weight (B) for dogs in the template group of Figure 1 and 18 similar dogs that comprised a test group after the dogs were classified as having a brachycephalic (template group, n = 10; test group, 6), mesaticephalic (template group, 11; test group, 6), or dolichocephalic (template group, 23; test group, 6) brain shape. For each plot, the horizontal line within the box represents the median; the lower and upper edges of the box delimit the 25th and 75th percentiles, respectively; and the whiskers delimit the 5th and 95th percentiles. Within each brain shape, results of 2-tailed Student t tests indicated that neither age nor body weight differed significantly (P > 0.05) between dogs in the template and test groups. See Figure 1 for remainder of key.
Citation: American Journal of Veterinary Research 77, 4; 10.2460/ajvr.77.4.395
Box-and-whisker plots of the data distributions for age (A) and body weight (B) for dogs in the template group of Figure 1 and 18 similar dogs that comprised a test group after the dogs were classified as having a brachycephalic (template group, n = 10; test group, 6), mesaticephalic (template group, 11; test group, 6), or dolichocephalic (template group, 23; test group, 6) brain shape. For each plot, the horizontal line within the box represents the median; the lower and upper edges of the box delimit the 25th and 75th percentiles, respectively; and the whiskers delimit the 5th and 95th percentiles. Within each brain shape, results of 2-tailed Student t tests indicated that neither age nor body weight differed significantly (P > 0.05) between dogs in the template and test groups. See Figure 1 for remainder of key.
Citation: American Journal of Veterinary Research 77, 4; 10.2460/ajvr.77.4.395
Accuracy of ABS
Results of the univariable logistic regression analyses indicated that age (P = 0.676) and body weight (P = 0.532) were not significantly associated with the Jaccard coefficient and were therefore not included in the multivariable analysis. The final multivariable model included segmentation method, brain shape, anatomic structure, and the interaction between segmentation method and brain shape. The model performed similarly for the internal brain structures (hippocampal formation and caudate nuclei), so the anatomic structure variable was collapsed into 2 categories (internal brain structures and brain). Sex was confounded by brain shape because the sex distribution varied among the brain shape categories. Given that brain shape was more strongly associated with the outcome of interest, sex was excluded from further analyses. There was a significant interaction between brain shape and segmentation method; model performance differed for ABS method C (manual brain extraction followed by application of the combined template for mesaticephalic and dolichocephalic brain shapes) and dogs with a dolichocephalic brain shape (Online Supplement available at http://avmajournals.avma.org/toc/ajvr/77/4).
Results of the final logistic regression model indicated that ABS method A (manual brain extraction and use of brain shape–specific template) was the most accurate of the 3 methods evaluated (ie, had the highest mean Jaccard coefficient), followed by method C and then method B. The mean Jaccard coefficient was 0.762 (95% CI, 0.716 to 0.803) for method A, 0.476 (95% CI, 0.402 to 0.557) for method B (automatic brain extraction followed by application of 1 of 3 brain shape–specific templates), and 0.630 (95% CI, 0.521 to 0.734) for method C. The marginal mean difference of the Jaccard coefficients was −0.283 (95% CI, −0.320 to −0.246; P < 0.001) between methods B and A, −0.128 (95% CI, −0.209 to −0.054; P < 0.001) between methods C and A, and 0.155 (95% CI, 0.060 to 0.241; P = 0.002) between methods C and B.
The Dice coefficients stratified by segmentation method and anatomic structure were summarized (Figure 3). The application of standard transformations to the Dice coefficients for each stratum was unsuccessful in normalizing the respective data distributions. The Dice coefficients ranged from 0.75 to 0.95 when manual brain extraction was used for ABS (methods A and C) and from 0.16 to 0.86 when automatic brain extraction was used for ABS (method B).
Tukey box-and-whisker plots of the Dice coefficients for 3 different methods (A [blue], B [orange], and C [green]) of MRI ABS for various anatomic structures of the canine brain. Method A involved manual brain extraction followed by application of 1 of 3 brain shape–specific (brachycephalic, mesaticephalic, or dolichocephalic) templates described in Figure 1. Method B involved automatic brain extraction followed by application of 1 of 3 brain shape–specific templates. Method C involved manual brain extraction followed by application of a combined template for mesaticephalic and dolichocephalic brain shapes. The Dice coefficient compares the extent of overlap between a source volume (method A, B, or C) and a target volume, or gold standard (manual segmentation), and is expressed as a proportion with a value of 1 indicating perfect overlap or unity between the volumes. The respective Dice coefficients were calculated by the application of each method to the MRI scans of 18 test dogs that met the same selection criteria as the dogs of the template group. For each plot, the horizontal line within the box represents the median; the lower and upper edges of the box delimit the 25th and 75th percentiles, respectively; the upper whisker delimits the largest value that is within the 75th percentile plus 1.5 times the interquartile range; the lower whisker delimits the smallest value or the 25th percentile minus 1.5 times the interquartile range, whichever is larger; and the squares represent outlier values. LtCN = Left caudate nucleus. LtHF = Left hippocampal formation. RtCN = Right caudate nucleus. RtHF = Right hippocampal formation. See Figure 1 for remainder of key.
Citation: American Journal of Veterinary Research 77, 4; 10.2460/ajvr.77.4.395
Tukey box-and-whisker plots of the Dice coefficients for 3 different methods (A [blue], B [orange], and C [green]) of MRI ABS for various anatomic structures of the canine brain. Method A involved manual brain extraction followed by application of 1 of 3 brain shape–specific (brachycephalic, mesaticephalic, or dolichocephalic) templates described in Figure 1. Method B involved automatic brain extraction followed by application of 1 of 3 brain shape–specific templates. Method C involved manual brain extraction followed by application of a combined template for mesaticephalic and dolichocephalic brain shapes. The Dice coefficient compares the extent of overlap between a source volume (method A, B, or C) and a target volume, or gold standard (manual segmentation), and is expressed as a proportion with a value of 1 indicating perfect overlap or unity between the volumes. The respective Dice coefficients were calculated by the application of each method to the MRI scans of 18 test dogs that met the same selection criteria as the dogs of the template group. For each plot, the horizontal line within the box represents the median; the lower and upper edges of the box delimit the 25th and 75th percentiles, respectively; the upper whisker delimits the largest value that is within the 75th percentile plus 1.5 times the interquartile range; the lower whisker delimits the smallest value or the 25th percentile minus 1.5 times the interquartile range, whichever is larger; and the squares represent outlier values. LtCN = Left caudate nucleus. LtHF = Left hippocampal formation. RtCN = Right caudate nucleus. RtHF = Right hippocampal formation. See Figure 1 for remainder of key.
Citation: American Journal of Veterinary Research 77, 4; 10.2460/ajvr.77.4.395
Tukey box-and-whisker plots of the Dice coefficients for 3 different methods (A [blue], B [orange], and C [green]) of MRI ABS for various anatomic structures of the canine brain. Method A involved manual brain extraction followed by application of 1 of 3 brain shape–specific (brachycephalic, mesaticephalic, or dolichocephalic) templates described in Figure 1. Method B involved automatic brain extraction followed by application of 1 of 3 brain shape–specific templates. Method C involved manual brain extraction followed by application of a combined template for mesaticephalic and dolichocephalic brain shapes. The Dice coefficient compares the extent of overlap between a source volume (method A, B, or C) and a target volume, or gold standard (manual segmentation), and is expressed as a proportion with a value of 1 indicating perfect overlap or unity between the volumes. The respective Dice coefficients were calculated by the application of each method to the MRI scans of 18 test dogs that met the same selection criteria as the dogs of the template group. For each plot, the horizontal line within the box represents the median; the lower and upper edges of the box delimit the 25th and 75th percentiles, respectively; the upper whisker delimits the largest value that is within the 75th percentile plus 1.5 times the interquartile range; the lower whisker delimits the smallest value or the 25th percentile minus 1.5 times the interquartile range, whichever is larger; and the squares represent outlier values. LtCN = Left caudate nucleus. LtHF = Left hippocampal formation. RtCN = Right caudate nucleus. RtHF = Right hippocampal formation. See Figure 1 for remainder of key.
Citation: American Journal of Veterinary Research 77, 4; 10.2460/ajvr.77.4.395
Repeatability of ABS
Method A was the most accurate ABS method evaluated. Therefore, the repeatability of method A of ABS and repeatability of manual segmentation were determined. The Dice coefficients were compared between the 2 methods (Table 1). Results indicated that ABS method A was extremely repeatable and, in fact, was more repeatable than manual segmentation.
Mean ± SD Dice coefficients calculated to assess the repeatability of MRI manual segmentation and ABS method A (manual brain extraction followed by application of 1 of 3 brain shape-specific [brachycephalic, mesaticephalic, or dolichocephalic] templates) when both were applied to a data set that consisted of 6 dogs from the test group (2 dogs with each brain shape) twice (rounds 1 and 2) with a minimum of 4 weeks between rounds.
| Segmentation method | ||
|---|---|---|
| Anatomic structure | Manual | ABS |
| Left hippocampal formation | 0.88 ± 0.02 | 1.00 ± 0.00 |
| Right hippocampal formation | 0.88 ± 0.01 | 1.00 ± 0.00 |
| Left caudate nucleus | 0.89 ± 0.02 | 1.00 ± 0.00 |
| Right caudate nucleus | 0.89 ± 0.03 | 1.00 ± 0.00 |
| Brain | 0.97 ± 0.01 | 1.00 ± 0.00 |
The Dice coefficient compares the extent of overlap between a source volume (from round 1) and a target volume (from round 2) and is expressed as a proportion with a value of 1 indicating perfect overlap or unity between the volumes. The templates were constructed from the MRI scans of 44 dogs without clinical signs of epilepsy and without MRI evidence of structural brain disease (template group). The respective Dice coefficients were calculated by the application of each segmentation method to the MRI scans for a subset of 6 of the 18 test dogs that met the same selection criteria as the dogs of the template group. Each round of segmentation involved the same data set.
Discussion
To our knowledge, the present study was the first to describe the construction of representative MRI atlases of the brain for dogs of various breeds, ages, brain shapes, and sex and assess the use of ABS in dogs with various brain shapes. Results indicated that ABS techniques were most accurate when the brain shape of the MRI atlas template used was similar to that of the individual being evaluated. Atlas-based segmentation methods that used manual brain extraction (methods A and C) were significantly more accurate than the ABS method that used automatic brain extraction (method B). Automatic brain extraction occasionally missed regions of the brain or included extraneous soft tissue, which resulted in the outliers observed for method B (Figure 3). Atlas-based segmentation method A was highly repeatable when it was applied to the same data set. This was not surprising because the application of ABS to the same data set is equivalent to replicating the same mathematical procedure. Manual segmentation was not as repeatable as ABS, most likely because of human error associated with tracing the margins of the brain structure being evaluated.
Other studies14–17 that involved the use of atlas-based registration for computer-assisted morphometry in dogs used templates derived from populations that were restricted in terms of breed, age, and sex. Prior to the present study, the most comprehensive MRI atlas of the canine brain was derived from 15 mesaticephalic dogs of various breeds.14 Although that atlas14 provides gyral and sulcal labels and identifies several internal brain structures, it does not include the olfactory lobes and was derived from 15 dogs that originated from familial lines with inherited retinal degeneration, 13 of which were affected from birth. Consequently, that atlas14 has limited applicability to nonmesaticephalic dogs. The atlases constructed for each of the 3 canine brain shapes (brachycephalic, mesaticephalic, and dolichocephalic) in the present study represent an improvement in the methodology for the generation of reliable and accurate brain atlases for dogs of various breeds, sizes, and brain shapes.
Identification of a method for consistent classification of dogs on the basis of brain morphology is problematic because of the marked variation in the brain shape of dogs. Several methods of classification were investigated prior to initiation of the present study. Breed-defined head shape and cephalic index do not always clearly distinguish between dogs with different brain shapes,19,26,27 and we believe the subjective evaluation of canine brain shape described in the present study is the best method for classifying brain shape. Subjective evaluation allows for the formation of a global opinion by taking into account a complex array of volumetric factors and spatial relationships on the basis of the Gestalt theory of visual perception.28 Those factors and relationships are not easily captured by a series of linear measurements or ratios. Subjective evaluation can, however, be difficult to reproduce, so brain length was used to classify dogs in the test group. Once an appropriate template is selected and ABS is performed, the segmented structure should always be visually inspected for accuracy of the segmentation; poor segmentation can be caused by inappropriate selection of a template or inherent distortion of brain anatomy (eg, hydrocephalus) that results in poor registration. Advancement of image registration and normalization tools may allow stable template construction for dogs with all brain shapes in the future; however, the validity of such a template for ABS will still need to be evaluated.
The open-source image analysis software program used in the present study contains a brain extraction tool for automatic extraction of human brains. This tool cannot accurately extract canine brains because the brain of a dog has a markedly different shape and contains substantially more extraneuronal tissue than does the brain of a human. In the present study, we used a template-based method for automatic brain extraction that was similar to that described by Datta et al,14 but automatic brain extraction was not as accurate as manual brain extraction and is not recommended. Manual brain extraction is the most labor intensive step in the ABS process and requires at least 30 min/brain to develop a customized brain mask. Development of a reliable and rapid method of automatic extraction of the canine brain warrants future research.
Following manual brain extraction, ABS of the 5 anatomic structures evaluated in the present study had measures of overlap that ranged from 0.75 to 0.95, which compared favorably with repeated measures of overlap for manual segmentation (gold standard) that ranged from 0.88 to 0.97 and indicated that ABS is a useful tool for segmentation of structures in large data sets. Additionally, the Dice coefficients for the caudate nuclei and hippocampus as determined by ABS following manual brain extraction (range, 0.6 to 0.9) were similar to the Dice coefficients for those structures (0.8), following the application of ABS in human medicine.18 In the present study, the univariable logistic regression coefficient for the brain was substantially higher than those for the hippocampal formation and caudate nuclei, which suggested that ABS was more accurate for the segmentation of structures with a small surface area-to-volume ratio than for structures with a large surface area-to-volume ratio. Univariable logistic regression results indicated that the ABS accuracy for the hippocampal formation did not differ from that for the caudate nuclei. This was a pleasing result because the caudate nuclei have a simple shape and are located in the rostral portion of the brain, whereas the hippocampus is located more caudally and has a complex coiled shape. Atlas-based segmentation of other neuroanatomic structures should be validated prior to the use of ABS for neuroanatomic research.
The MRI atlases constructed for the present study were derived from dogs of various breeds and ages in an attempt to produce unbiased templates representative of the general dog population. However, it is important to note that Greyhounds were overrepresented in the dolichocephalic template, which biased that template for that breed. That bias might explain the interaction between brain shape and segmentation method observed during multivariable logistic regression in which ABS with the combined (mesaticephalic and dolichocephalic) template performed better for dogs with a dolichocephalic brain shape. Whenever possible, MRI atlas templates constructed from a sample population that is similar to the test group (eg, age-, breed-, and sex-matched controls) should be used for ABS because those templates should produce more accurate segmentation results than the use of generalized templates. Unfortunately, the availability of MRI scans of clinically normal canine brains is limited.
A limitation of the MRI atlases generated by the methods described in the present study was that, although no obvious anatomic abnormality was present on the MRI scans of any of the 44 dogs in the template group, many of those dogs had neurologic signs such as vestibular disease (n = 12 [27%]) or behavioral changes (4 [9%]). The effect of undetected structural pathology on the validity of the MRI atlas templates constructed is unknown and represents the greatest limitation for the future clinical and research applications of those templates. Nonetheless, the present report provided an important description of the application of the ABS technique in dogs. Investigators interested in implementing ABS as a research tool are encouraged to carefully consider the selection of control dogs when constructing MRI atlases of the brain. Also, although the MRI scans used in the present study were acquired by the same MRI machine, 2 different coils were used and the matrix size varied, depending on size of the dog. The effect of different machines, coils, and variables such as matrix size on the quality of image registration and accuracy of ABS is unknown. Prospective studies in which those variables are controlled are warranted; however, standardization of those variables in clinical settings may not be possible.
In the present study, we successfully developed 3 MRI atlases for the canine brain that were specific for brachycephalic, mesaticephalic, and dolichocephalic brain shapes, and we documented that the application of ABS with a brain shape–specific template following manual brain extraction was a valid technique for computer-assisted morphometry in dogs that is both accurate and repeatable. Atlas-based segmentation may be beneficial for canine neuroimaging research, especially for the investigation of diseases that are currently classified as idiopathic or those with occult underlying etiologies such as epilepsy, canine cognitive dysfunction syndrome, or behavior disorders. Furthermore, ABS might provide a basis for investigating the correlation between breed or behavior and neurologic structures. Atlas-based segmentation can be performed more rapidly than manual segmentation, and incorporation of ABS into automatic scripts and software with a user-friendly graphic interface will likely increase the use of volumetry in clinical neuroimaging.
ABBREVIATIONS
| ABS | Atlas-based segmentation |
| CI | Confidence interval |
Footnotes
Signa, GE Healthcare, Little Chalfont, Buckinghamshire, England.
Omniscan, GE Healthcare, Little Chalfont, Buckinghamshire, England.
Dell PowerEdge C6145 with 256Gb of RAM using the Sun Grid Engine (SGE) Oracle, Oracle and Sun Microsystems, Redwood Shores, Calif.
FSL, version 5.0.7. FMRIB Analysis Group, Oxford, England.
ITK-SNAP, versions 2.0.0 and 3.0.0, Yushkevich P, Penn Image Computing and Science Laboratory, Department of Radiology, University of Pennsylvania, Philadelphia, Penn, and Gerig G, Scientific Computing and Imaging Institute, University of Utah, Salt Lake City, Utah.
Advanced Normalization Tools (ANTs), version 1.9.4, Penn Image Computing and Science Laboratory, Philadelphia, Penn.
GraphPad Prism, version 6.0, GraphPad Software Inc, La Jolla, Calif.
Stata/MP, version 13.1, StataCorp LP, College Station, Tex.
SPSS, version 22, IBM Corp, Armonk, NY.
References
1. Ashburner J, Csernansky JG, Davatzikos C, et al. Computer-assisted imaging to assess brain structure in healthy and diseased brains. Lancet Neurol 2003; 2: 79–88.
2. Bernasconi A, Bernasconi N, Bernhardt BC, et al. Advances in MRI for “cryptogenic” epilepsies. Nat Rev Neurol 2011; 7: 99–108.
3. Hudson JA, Simpson ST, Buxton DF, et al. Ultrasonographic diagnosis of canine hydrocephalus. Vet Radiol Ultrasound 1990; 31: 50–58.
4. Carrera I, Dennis R, Mellor DJ, et al. Use of magnetic resonance imaging for morphometric analysis of the caudal cranial fossa in Cavalier King Charles Spaniels. Am J Vet Res 2009; 70: 340–345.
5. Henke D, Böttcher P, Doherr MG, et al. Computer-assisted magnetic resonance imaging brain morphometry in American Staffordshire Terriers with cerebellar cortical degeneration. J Vet Intern Med 2008; 22: 969–975.
6. Thames RA, Robertson ID, Flegel T, et al. Development of a morphometric magnetic resonance image parameter suitable for distinguishing between normal dogs and dogs with cerebellar atrophy. Vet Radiol Ultrasound 2010; 51: 246–253.
7. Roberts T, McGreevy P, Valenzuela M. Human induced rotation and reorganization of the brain of domestic dogs. PLoS ONE 2010;5:e11946.
8. Tapp PD, Siwak CT, Gao FQ, et al. Frontal lobe volume, function, and beta-amyloid pathology in a canine model of aging. J Neurosci 2004; 24: 8205–8213.
9. Su MY, Tapp PD, Vu L, et al. A longitudinal study of brain mor-phometrics using serial magnetic resonance imaging analysis in a canine model of aging. Prog Neuropsychopharmacol Biol Psychiatry 2005; 29: 389–397.
10. Driver CJ, Rusbridge C, Cross HR, et al. Relationship of brain parenchyma within the caudal cranial fossa and ventricle size to syringomyelia in Cavalier King Charles Spaniels. J Small Anim Pract 2010; 51: 382–386.
11. Kuwabara T, Hasegawa D, Kobayashi M, et al. Clinical magnetic resonance volumetry of the hippocampus in 58 epileptic dogs. Vet Radiol Ultrasound 2010; 51: 485–490.
12. Milne ME, Anderson GA, Chow KE, et al. Description of technique and lower reference limit for magnetic resonance imaging of hippocampal volumetry in dogs. Am J Vet Res 2013; 74: 224–231.
13. Schmidt MJ, Amort KH, Failing K, et al. Comparison of the endocranial- and brain volumes in brachycephalic dogs, mesaticephalic dogs and Cavalier King Charles Spaniels in relation to their body weight. Acta Vet Scand 2014;56:30.
14. Datta R, Lee J, Duda J, et al. A digital atlas of the dog brain. PloS ONE 2012;7:e52140.
15. Jacqmot O, Van Thielen B, Fierens Y. Diffusion tensor imaging of white matter tracts in the dog brain. Anat Rec (Hoboken) 2013; 296: 340–349.
16. Tapp PD, Head K, Head E, et al. Application of an automated voxel-based morphometry technique to assess regional gray and white matter brain atrophy in a canine model of aging. Neuroimage 2006; 29: 234–244.
17. Jia H, Pustovyy OM, Waggoner P, et al. Functional MRI of the olfactory system in conscious dogs. PLoS ONE 2014;9:e86362.
18. Cabezas M, Oliver A, Lladó X, et al. A review of atlas-based segmentation for magnetic resonance brain images. Comput Methods Programs Biomed 2011; 104:e158–e177.
19. Evans HE, Miller ME. Miller's anatomy of the dog. 3rd ed. Philadelphia: WB Saunders Co, 1993.
20. Jenkinson M, Beckmann CF, Behrens TEJ, et al. FSL. Neuroimage 2012; 62: 782–790.
21. Yushkevich PA, Piven J, Hazlett HC, et al. User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability. Neuroimage 2006; 31: 1116–1128.
22. Avants BB, Tustison NJ, Song G, et al. A reproducible evaluation of ANTs similarity metric performance in brain image registration. Neuroimage 2011; 54: 2033–2044.
23. Avants BB, Epstein CL, Grossman M, et al. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Med Image Anal 2008; 12: 26–41.
24. Avants B, Gee JC. Geodesic estimation for large deformation anatomical shape averaging and interpolation. Neuroimage 2004; 23(suppl 1): S139–S150.
25. Tustison NJ, Gee JC. Introducing Dice, Jaccard, and other label overlap measures to ITK. Insight J 2009;Jul-Dec:707.
26. Onar V, Ozcan S, Pazvant G. Skull typology of adult male Kangal dogs. Anat Histol Embryol 2001; 30: 41–48.
27. Schmidt MJ, Neumann AC, Amort KH, et al. Cephalometric measurements and determination of general skull type of Cavalier King Charles Spaniels. Vet Radiol Ultrasound 2011; 52: 436–440.
28. Navon D. Forest before trees: the precedence of global features in visual perception. Cognit Psychol 1977; 9: 353–383.
