Caution: shortcomings of traditional segmentation methods from magnetic resonance imaging brain scans intended for 3-dimensional surface modelling in children with pathology.
Anith ChackoSean SchoemanShyam Sunder B VenkatakrishnaSamuel BoltonAndrew I U ShearnSavvas AndronikouPublished in: Pediatric radiology (2023)
This technical innovation assesses the adaptability of some common automated segmentation tools on abnormal pediatric magnetic resonance (MR) brain scans. We categorized 35 MR scans by pathologic features: (1) "normal"; (2) "atrophy"; (3) "cavity"; (4) "other." The following three tools, (1) Computational Anatomy Toolbox version 12 (CAT12); (2) Statistical Parametic Mapping version 12 (SPM12); and (3) MRTool, were tested on each scan-with default and adjusted settings. Success was determined by radiologist consensus on the surface accuracy. Automated segmentation failed in scans demonstrating severe surface brain pathology. Segmentation of the "cavity" group was ineffective, with success rates of 23.1% (CAT12), 69.2% (SPM12) and 46.2% (MRTool), even with refined settings and manual edits. Further investigation is required to improve this workflow and automated segmentation methodology for complex surface pathology.
Keyphrases
- deep learning
- contrast enhanced
- computed tomography
- convolutional neural network
- magnetic resonance
- magnetic resonance imaging
- resting state
- machine learning
- white matter
- functional connectivity
- dual energy
- high throughput
- psychometric properties
- squamous cell carcinoma
- early onset
- young adults
- diffusion weighted imaging
- mass spectrometry
- high density
- blood brain barrier
- rectal cancer
- electronic health record