Comparison of manual and automated ventricle segmentation in the maternal immune stimulation rat model of schizophrenia

Maternal immune stimulation (MIS) is strongly implicated in the etiology of neuropsychiatric disorders. Magnetic resonance imaging (MRI) studies provide evidence for brain structural abnormalities in rodents following prenatal exposure to MIS. Reported volumetric changes in adult MIS offspring comprise among others larger ventricular volumes, consistent with alterations found in patients with schizophrenia. Linking rodent models of MIS with non-invasive small animal neuroimaging modalities thus represents a powerful tool for the investigation of structural endophenotypes. Traditionally manual segmentation of regions-of-interest, which is laborious and prone to low intra- and inter-rater reliability, was employed for data analysis. Recently automated analysis platforms in rodent disease models are emerging. However, none of these has been found to reliably detect ventricular volume changes in MIS nor directly compared manual and automated data analysis strategies. The present study was thus conducted to establish an automated, structural analysis method focused on lateral ventricle segmentation. It was applied to ex-vivo rat brain MRI scans. Performance was validated for phenotype induction following MIS and preventive treatment data and compared to manual segmentation. In conclusion, we present an automated analysis platform to investigate ventricular volume alterations in rodent models thereby encouraging their preclinical use in the search for new urgently needed treatments.


22
Exploring new avenues in psychiatric research has the goal of combing treatment effectiveness 23 with tolerability, often with the idea of a preventive treatment strategy. One condition in order to achieve 24 this high-set goal is the availability of cross-species biomarkers. Good parameters should be easily

37
Brain volumes can be easily measured with magnetic resonance imaging (MRI), a non-invasive 38 imaging method available with a good soft-tissue contrast and good spatial resolution (50 -100 µm) 39 (16,17). It utilizes longitudinal magnetization signals to create contrast images whose intensity range 40 indicate structural boundaries and tissue properties in order to further make informed decisions.

152
The N4BiasFieldCorrection algorithm (N4ITK), an optimized version of the N3 that 153 implements a robust B-spline approximation routine and a modified optimization scheme (36), was used 154 to correct intensity non-uniformity within images. 50 iterations over three levels with a convergence 155 threshold of 1e-6, full width at half maximum (deconvolution kernel) of 0.15mm were employed with 156 default parameters to provide optimum results for the rat subject MR images. For brain extraction the 157 python command line SkullStrip (37) was used for the semi-automated registration brain extraction in this 158 project. A reference stripped image with its brain mask (manually delineated) and the brain image was 159 used to generate a stripped fixed image mask that was then binarized. Image multiplication operation 160 was then performed with this mask and the original subject image (see S1 Appendix for details).

189
MR images from 42 animals (n=21 saline, n=20 poly) were analyzed using the established 190 analysis platform, labels were visually quality controlled but not corrected and ventricular volumes 191 estimated. As expected, estimates were observed to be lower in saline 4.198 ± 0.247 mm³ compared to 192 poly I:C 4.434 ± 0.279 mm³ offspring, respectively (Fig. 2a,b). The difference was found to be 193 significant T(39)= -2.858 p= .007 (Fig 2c). Due to the enormous labor intensity only half of the sample 194 was also evaluated manually (n=12 saline, n=12 poly; for example of manual segmented image see S1 For further performance validation, imaging data from (8), which was published employing 211 manual delineation, was now re-analyzed with the newly established automated analysis platform.

212
Following visual quality control, ventricle labels were manually corrected for 15% of subjects and LV 213 volume statistics obtained (Fig. 3)

253
Our analysis included ex-vivo MRI data from a study of tDCS application to prevent 254 manifestation of schizophrenia associated deficits in MIA rats (8). tDCS is a brain stimulation method 255 were low electric positive (anodal) or negative (cathodal) current is applied to an area of the brain for 256 the depolarization or hyperpolarization of neurons towards facilitatory or inhibitory behavioral effects 257 respectively (58). Medial prefrontal cortex (mPFC) stimulation, based on established protocols (59) and 258 applied during adolescence, was reported to restrict the emergence of schizophrenia-related behaviors 259 as well as volumetric changes. The previously, manually segmented data sets were now reanalyzed to 260 confirm phenotype and treatment effects at adulthood. We were able to replicate the results with our 261 automated analysis platform, showing a significant difference in lateral ventricle volumes with lower 262 volumes in saline offspring compared to poly I:C offspring. In addition, the automated analysis platform 263 confirmed lower lateral ventricle sizes for poly I:C offspring treated with anodal stimulation but not for 264 poly I:C offspring treated with cathodal stimulation.

265
Manual segmentation is still frequently used to achieve volumetric measures from neuroimaging 266 techniques due to its robustness, although it has certain disadvantages. It is very time consuming, taking 267 2-3 hours per animal for a practiced expert, and therefore labor-intensive. Variance in the results is high, 268 with inter-and intra-rater biases, and the need for a better alternative is formidable. The automated 269 method as shown in this paper is accurate, allows repetitiveness and is much less time-consuming.

270
However, automated segmentation still comes with various challenges. They include often low image 271 resolution of rodent brains (smaller voxel sizes in rodents compared to humans due to their brain size), 272 physiological noise and signal-to-noise ratio (57,60 and the preparation of data to reduce a potential systematic bias. In the used registration method, a brain 286 atlas with its corresponding brain mask is used for brain extraction. Therefore, the brain extraction is dependent on the availability of an atlas, its mask and the parameters of registration. The rats' size, age 288 and the similarity between the reference atlas image and the target image of the dataset must be 289 considered. To aid this, a reference subject can be chosen from the study cohort, manually segmented 290 and used for the automated registration brain extraction methods. In our approach, the SkullStrip (37) 291 command tool, a robust automated brain extraction tool, was used, which is dependent on the image 292 resolution. This tool was further optimized by binarizing the skull stripped output for each subject and 293 applying it to the original image to get the best "brain-only" volume image for each subject. These 294 additional steps resulted in a higher correlation in visual inspection, total brain volume and similarity 295 measure to the manual method. These efforts may be partially responsible for achieving results in line 296 with previous work on characterizing volumetric changes in lateral ventricles with the presented 297 platform contrary to the findings of other semi-automated segmentation approaches (26).

298
Limitations of the presented analysis platform include the still labor-intensive manual extracting 299 of the reference subject and its brain mask used with the optimized skull-strip method. This step, though 300 advisable to remove bias resulting from differences in strain and age when needed, is only carried out 301 once. A within-study template is highly rocommended. Second the image contrast, significant for brain 302 segmentation was low in our data. There is a need for a standard segmentation protocol detailing the 303 process to delineate rodent brains for imaging modalities other than histology. Anderson