Purpose or Case Report: To assess the utility and adaptability of some widely used automated segmentation methods when applied to abnormal pediatric magnetic resonance imaging (MRI) brain scans. Segmentation is an essential component of the workflow when building 3D anatomical models of abnormal pediatric brains to demonstrate surface pathology.
Methods & Materials: Institutional review board approval was obtained. 35 abnormal pediatric MRI brain scans were included from a larger population of children with cerebral palsy who received delayed neuroimaging for suspected neonatal hypoxic ischemic injury. They were reviewed by two attending pediatric neuroradiologists and distributed into 4 groups based on pathologic features: 1) Normal; 2) Atrophic; 3) Cavity; 4) Other/ non-cortical. Three automated segmentation tools, all accessible via the Matlab computing environment were used: 1) Computational Anatomy Toolbox version 12 (CAT12); 2) Statistical Parametic Mapping version 12 (SPM12); 3) MRTool. These methods use a combination of tissue probability mapping (TPM) and voxel-based morphometry (VBM) segmentation methods. Segmentation was performed by all 3 modalities on each patient, first with default settings, then with adjusted settings. Successful segmentation was determined by radiologist consensus on the generation of a surface mesh accurately reflecting the source scans that then could be successfully printed.
Results: Initial segmentation with default settings produced 20/35 (57.1%) usable models. Further refinement of various criteria in each modality yielded an increase to successful segmentation in 31/35 (88.6%) models. The cavity group was the most difficult to segment under default settings with overall low initial success rates of CAT12 15.4%, SPM12 23.1%, and MRtool 15.4%. Refined settings and manual edits improved these rates to 23.1%, 69.2%, and 46.2%, respectively.
Conclusions: Segmentation, when using SPM12, CAT12 and MRtool was ineffective as applied to categories of abnormal brain scans demonstrating severe surface pathology - atrophy and especially cavities with thin overlying cortex. Segmentation is a vital component in the workflow when creating 3D anatomical models. The tested methods are not suitable, even with manual augmentation when attempting to segment pediatric brains with severe pathology. Further investigation is required to improve this workflow and create automated segmentation methods that maintain anatomical accuracy when applied to complex surface pathology.