Fully automatic adaptive meshing based segmentation of the ventricular system for augmented reality visualization and navigation

: OBJECTIVE Effective image segmentation of cerebral structures is fundamental to 3-D techniques such as augmented reality (AR). To be clinically viable, segmentation algorithms should be fully-automatic and easily integrated in existing digital infrastructure. We created a fully automatic adaptive-meshing-based segmentation system for T1-weighted magnetic resonance images (MRI) to automatically segment the complete ventricular system, running in a cloud-based environment that can be accessed on an AR device. This study aims to assess the accuracy and segmentation time of the system by comparing it to a manually segmented ground truth dataset. METHODS A ground truth (GT) dataset of 46 contrast-enhanced and non-contrast-enhanced T1-weighted MRI scans was manually segmented. These scans were also uploaded to our system to create a machine-segmented (MS) dataset. The GT data was compared to the MS data using the Sørensen-Dice similarity coeﬀicient (DSC) and 95% Hausdorff distance (HD 95 ) to determine segmentation accuracy. Furthermore, segmentation times for all GT and MS seg-mentations were measured. RESULTS Automatic segmentation was successful for 45 (98%) of 46 cases. Mean DSC score was 0.83 (SD = 0.08) and mean HD 95 was 19.06 mm (SD=11.20). Segmentation time was significantly longer for the GT group (mean=14405 s, SD=7089) when compared to the MS group (mean=1275 s, SD=714) with a mean difference of 13130 s (95% CI [10130, 16130]). CONCLUSIONS The described adaptive meshing based segmentation algorithm provides accurate and time-eﬀicient automatic segmentation of the ventricular system from T1 MRI scans and direct visualization

Accurate segmentation of the ventricular system could provide certain advantages when 51 supporting diagnostics, pre-operative planning or intraoperative navigation. For diagnostics, 52 automated volumetric assessment of each subpart of the ventricles (lateral, third, fourth) would 53 be preferable over subjective 2-dimensional measuring over random slices, which is often the 54 current standard. For pre-operative planning, difficult anatomical variations (such as in some 55 congenital cases) could be better prepared and understood if a surgeon is provided with a 56 detailed 3-D model of the anatomical region-of-interest, especially when combined with AR or 57 3-D printing. Moreover, 3-D models are an asset in the education of patients, residents or 58 medical students, as 3-D images are more effective to develop spatial understanding than 2-D 59 images 1-7 . For intraoperative navigation, the model of the ventricle or the trajectory to the chosen 60 entry site of the ventricle could be used to guide a surgeon during ventricular drain placement. 61 J o u r n a l P r e -p r o o f Cambridge, USA). For this dataset, the ventricular system of each MRI scan was manually 100 painted using a "brush" tool on a slice-by-slice basis ( Figure 1A). After segmenting all slices of a 101 scan, the volumetric 3-D segmentation was exported in Neuroimaging Informatics Technology 102 Initiative (NIfTI) file format. The segmentation was also converted to a 3-D surface model and 103 exported in Stereolithography (STL) file format. Segmentations were performed by one of 4 104 authors (JvD, MA, MK, VvdK). To optimize the accuracy of each segmentation, a written 105 guideline (Supplementary material Appendix 1) was established that dictated exactly which 106 anatomical structures to include in the segmentation. Furthermore, difficult cases were discussed 107 collectively in a weekly meeting. As final step, all cases were evaluated and corrected slice by 108 slice by the senior author (TvD). If the segmentation was deemed of sufficient quality, the 109 segmentation was included for quantitative analysis. 110 Automatic segmentation 111 The segmentation system was embedded within an online cloud environment that stored all 112 segmentation data (Augsafe, Augmedit, Naarden, The Netherlands). This environment was built, 113 secured and hosted using a cloud computing service (Azure, Microsoft, Redmond, USA). To use 114 the cloud environment and segmentation algorithm, a web-based user interface (UI) was 115 developed which could be accessed using a personal computer or an AR-HMD (Hololens, 116 Microsoft, Redmond, USA) ( Figure 2). Using this interface, MR images were uploaded in 117 Digital Imaging and Communications in Medicine (DICOM) format and subsequently 118 automatically segmented on an external server which supported simultaneous computation of 119 multiple segmentations. 120 The automatic segmentation was performed using an expanding mesh algorithm (Disior, 121 Helsinki, Finland). This algorithm uses 3-D adaptive spheres that expand and capture the 122 radiological boundaries of various tissues, including the ventricles. For each scan, this algorithm 123 is tailored to the specific patient by pre-processing using region, threshold and histogram based 124 segmentation methods. Previously, this method has been used in orbital volume calculations 22-27 . 125 Each included scan was sequentially segmented using this system. After segmentation, the mesh 126 was automatically simplified to streamline 3-D rendering. All segmentation steps were 127 performed without any user input. 128 The web-based interface was used to access all segmentation results in a 3-D viewer ( Figure 1B

178
In total, 46 scans were included, whereof 45 (98%) scans were successfully segmented and 179 computation failed for 1 (2%) case, which failed due to a data reading error. Clinical applications 266 The currently described segmentation system could potentially be used to support AR 267 navigation, particularly to guide a ventricular shunt placement 8,9 . In this case, AR 268 neuronavigation methods require accurate anatomical 3-D models to guide the intraventricular 269 part of the trajectory. Furthermore, segmentation should also be reasonably fast to prevent 270 disruption of clinical workflows. As the interest in AR neuronavigation has increased over the 271 last years, segmentation techniques that are tailored to this modality are necessary. As our system 272 operates completely automatically, is reasonably fast and accurate, can be connected to a 273 hospital's PACS system and can be accessed using AR-HMDs, we believe that it could fulfill 274 this role. Furthermore, the system could benefit diagnostics and follow-up of cases of 275 hydrocephalus. 276

Future improvements 277
The currently described system has several potential improvements. First, to decrease the 278 incidence of data-leaks, we will attempt to make the algorithm more sensitive to very thin 279 boundaries between cavities. Very high resolution scans of 400 by 400 pixels or more, which 280 may be more common in the future, would also decrease the frequency of leaks. Furthermore, we 281 aim to automatically recognize intraventricular blood, tumors and the choroid plexus, which 282 could also increase the accuracy of the ventricles segmentation by using subtraction. Lastly, we 283 have planned to add a post-processing environment, in which the surgeon can manually improve 284 the segmentation using boundary boxes, erasers and trajectory lines. 285 In the future, we will incorporate segmentation of T2 weighted images. This could potentially 286 increase segmentation accuracy, as cerebrospinal fluid induces a high signal on this sequence. 287 Previous studies using T2 weighted images have reported promising results 14,15      J o u r n a l P r e -p r o o f Introduction The ventricular system is an anatomically complex 3-dimensional structure that has a distinctive shape on an MRI scan. As the structure is filled with cerebrospinal fluid (CSF), the system is either completely black (T1) or white (T2) on an MRI scan, and forms a continuous cavity. The ventricular system consists of 4 segments, whereas the lateral (1 st and 2 nd ) ventricles connect to the 3d ventricle using the foramen of Monro. The 3d ventricle connects to the 4 th ventricle using the aqueduct of Sylvius.
Although the structure can be recognized by the human eye on MRI scans, no commercially available systems for the fully automatic segmentation of the structure have been developed. This project aims to develop a fully automatic back-end function for medical segmentation software that can segment, among other brain structures, the complete ventricular system. To scrutinize the accuracy of this algorithm, an extensive benchmark set of optimally segmented ventricular system 3D data is needed. This can then be used to compare volumes and surface geometries of automatic segmentation to, functioning as a 'golden standard'. As 3D segmentation of MRI scans is a relatively new field of study, this dataset is currently not available. For this reason, we will manually segment the ventricular system from 46 MRI scans as a kick-off for this study, to create our own extensive benchmark set.
Although manual segmentation is relatively straightforward, it is very important to carry out correctly for a proper analysis. Furthermore, it is imperative to establish a set of rules to keep the segmentation method homogenous between researchers. This document will function as a guideline for the standard method of manual segmentation.

Instruction
Software/hardware requirements: