Validation of an automated segmentation algorithm for the volumetric analysis of perihematomal edema in patients with spontaneous intracerebral hemorrhage

Primary Authors: Natasha Ironside, MBChB1; Simukayi Mutasa, MD2; Sachin Jambawalikar, PhD3; Jan Claassen, MD4; David Roh, MD4; Angela Lignelli, MD2; Stephen Mayer, MD5; Edward Sander Connolly, Jr., MD1

1Department of Neurological Surgery, Columbia University Irving Medical Center, New York, New York
2Department of Radiology, Columbia University Irving Medical Center, New York, New York
3Department of Medical Physics, Columbia University Irving Medical Center, New York, New York
4Department of Neurology, Columbia University Irving Medical Center, New York, New York
5Department of Neurology, Henry Ford Health System, Detroit, Michigan

Intracerebral hemorrhage (ICH), which accounts for up to 15% of all strokes, carries the highest rate of stroke-related death and long-term disability.1 Perihematomal edema contributes to the mass effect of the initial bleed and to ischemic toxicity within the surrounding parenchyma.2,3 Both the rate and extent of edema growth have previously been shown to independently affect mortality and functional outcomes.4-7
Several limitations to the volumetric analysis of perihematomal edema have been reported. While the signal intensity difference seen on T2-weighted magnetic resonance imaging (MRI) effectively delineates perihematomal edema from the surrounding parenchyma, MRI is not routinely obtained in unstable ICH patients or as a serial imaging modality.8 On computed tomography (CT), perihematomal edema manifests as a hypodensity that poses challenges to semi-automated threshold-based segmentation, due to its similarity in appearance to cerebrospinal fluid (CSF) and microangiopathy.9,10 Manual segmentation, which is both time consuming and cumbersome to accomplish, has high reported rates of intra and inter-user variability.10
The application of convolutional neural networks, a deep learning method, can overcome the aforementioned limitations by developing an algorithm capable of learning the features of image interpretation without any direct human intervention during the training process.11 Convolutional neural networks have been successfully trained to automatically segment perihematomal edema in our single center cohort of ICH patients (unpublished data). The external validation of such a tool may permit the systematic investigation of perihematomal edema in prospective, multi-center studies and provide reliable criteria for assessing the efficacy of novel treatments.

Convolutional neural networks for the volumetric analysis of perihematomal edema perform equal to or better than both manual and semi-automated methods in an externally derived patient cohort.

1. Perihematomal edema volume

1. Segmentation time
2. Inter-observer reliability
3. Intra-observer reliability

Patient characteristics and image selection
Patients will be derived from the VISTA-ICH database. Baseline demographic, clinical and radiologic characteristics will be recorded. A random sample of 500 CT scans will be de-identified and transferred in DICOM format to a central workstation.

Image processing
For automated segmentation, the DICOMs will be converted to NumPy array format before loading onto the fully convolutional neural network architecture. The NumPy array permits computationally efficient multi-dimensional data storage. For semi-automated segmentation, the DICOMs will be converted into Neuroimaging Informatics Technology Initiative (NIfTI) format using a DCM-to-NIfTI conversion tool before loading on to Analyze 12.0 (Biomedical Imaging Resource, Mayo Clinic). The NIfTI allows for individual DICOM images to be saved and loaded from one single file. For manual segmentation, the DICOMs will be loaded directly onto 3D Slicer 4.8 (National Institutes of Health, Bethesda, MD).

Segmentation
For automated segmentation, a fully convolutional neural network architecture, constructed in TensorFlow, will be implemented by performing a series of up sampling convolutional transpose operators on the deepest network layers.12 The resultant dense classification matrix will be equal in dimension to the original image size for each forward pass.13 Symmetric contracting and expanding topology will be employed to efficiently combine low- and high- level features.

For semi-automated segmentation, the method reported by Volbers et al. will be used.10 A limit boundary will be placed around the ICH and edema complex, following which a seed point will be placed within the region of interest (ROI). The lower Hounsfield Unit (HU) limit will be fixed at 5, with the upper boundary manually adjusted to a maximum of 33-HU using visual inspection and comparison to the contralateral hemisphere for identification of hypodensity attributable to microangiopathy.
For manual segmentation, the perifocal hypodense area surrounding the hemorrhage will be traced on each slice using visual inspection and comparison to the contralateral hemisphere for identification of hypodensity attributable to microangiopathy.

Reliability and workflow
All manual and semi-automated segmentation steps will be performed independently by eight neuro-imaging experts (neurology and neuroradiology fellows). Four observers will be assigned to semi-automated and four assigned to manual segmentations. Each observer will be blinded to the results obtained by other observers. For evaluation of intra-observer reliability, repeat segmentations will be performed for each of the three methods, after a minimal interval of 7 days. The time taken to perform segmentation, at each instance, will be recorded for each of the three methods.

Volumetric analysis
Edema volumes will be calculated by multiplying the number of segmented voxels by the distance between each voxel in the x, y and z dimensions.9,10 The z dimension will be calculated by estimating the gantry tilt adjusted voxel depth for each slice.9

Statistical analysis
Baseline demographic, clinical and radiographic characteristics among the patient cohort will be summarized. For each of the three segmentation methods, edema volumes and segmentation times will be compared using one-way analysis of variance (ANOVA) with Tukey’s pairwise comparisons to evaluate for differences between the groups. Inter- and intra-observer reliability will be evaluated using the intraclass correlation coefficient. Statistical significance will be defined as p<0.05, and all tests will be two-tailed.

1. Sacco S, Marini C, Toni D, Olivieri L, Carolei A. Incidence and 10-year survival of intracerebral hemorrhage in a population-based registry. Stroke 2009; 40(2): 394-9.
2. Zheng H, Chen C, Zhang J, Hu Z. Mechanism and Therapy of Brain Edema after Intracerebral Hemorrhage. Cerebrovasc Dis 2016; 42(3-4): 155-69.
3. Keep RF, Hua Y, Xi G. Intracerebral haemorrhage: mechanisms of injury and therapeutic targets. Lancet Neurol 2012; 11(8): 720-31.
4. Murthy SB, Urday S, Beslow LA, et al. Rate of perihaematomal oedema expansion is associated with poor clinical outcomes in intracerebral haemorrhage. J Neurol Neurosurg Psychiatry 2016; 87(11): 1169-73.
5. Urday S, Beslow LA, Dai F, et al. Rate of Perihematomal Edema Expansion Predicts Outcome After Intracerebral Hemorrhage. Crit Care Med 2016; 44(4): 790-7.
6. Volbers B, Giede-Jeppe A, Gerner ST, et al. Peak perihemorrhagic edema correlates with functional outcome in intracerebral hemorrhage. Neurology 2018; 90(12): e1005-e12.
7. Wu TY, Sharma G, Strbian D, et al. Natural History of Perihematomal Edema and Impact on Outcome After Intracerebral Hemorrhage. Stroke 2017; 48(4): 873-9.
8. Carhuapoma JR, Hanley DF, Banerjee M, Beauchamp NJ. Brain edema after human cerebral hemorrhage: a magnetic resonance imaging volumetric analysis. J Neurosurg Anesthesiol 2003; 15(3): 230-3.
9. Wu TY, Sobowale O, Hurford R, et al. Software output from semi-automated planimetry can underestimate intracerebral haemorrhage and peri-haematomal oedema volumes by up to 41. Neuroradiology 2016; 58(9): 867-76.
10. Volbers B, Staykov D, Wagner I, et al. Semi-automatic volumetric assessment of perihemorrhagic edema with computed tomography. Eur J Neurol 2011; 18(11): 1323-8.
11. Chartrand G, Cheng PM, Vorontsov E, et al. Deep Learning: A Primer for Radiologists. Radiographics 2017; 37(7): 2113-31.
12. Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2015; 39(4): 640-51.
13. Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation. International Conference on Medical Image Computing and Computer-Assisted Intervention - Lecture Notes in Computer Science. Springer 2015; 9351:234-41.