A comparison of automated lesion segmentation approaches for chronic stroke T1‐weighted MRI data

• Published in: Human brain mapping Vol. 40; no. 16; pp. 4669 - 4685
• Main Authors: Ito, Kaori L., Kim, Hosung, Liew, Sook‐Lei
• Format: Journal Article
• Language: English
• Published: Hoboken, USA John Wiley & Sons, Inc 01.11.2019
• Subjects: Algorithms; Automation; Bayes Theorem; Bayesian analysis; big data; Brain - diagnostic imaging; Brain stem; Cerebellum; Chronic Disease; chronic stroke; Cluster Analysis; Data analysis; Datasets; False Negative Reactions; Functional anatomy; Humans; Image Interpretation, Computer-Assisted; Image processing; Image Processing, Computer-Assisted; Image segmentation; lesion segmentation; Lesions; Literature reviews; Magnetic Resonance Imaging - methods; Medical imaging; MRI; Neuroimaging; Neurology; Normal Distribution; Rehabilitation; Stroke; Stroke - diagnostic imaging; Structure-function relationships; Territory; lesion segmentation; chronic stroke; big data; MRI; stroke
• ISSN: 1065-9471; 1097-0193; 1097-0193
• DOI: 10.1002/hbm.24729

Accurate stroke lesion segmentation is a critical step in the neuroimaging processing pipeline for assessing the relationship between poststroke brain structure, function, and behavior. Many multimodal segmentation algorithms have been developed for acute stroke neuroimaging, yet few algorithms are effective with only a single T1‐weighted (T1w) anatomical MRI. This is a critical gap because multimodal MRI is not commonly available due to time and cost constraints in the stroke rehabilitation setting. Although several attempts to automate the segmentation of chronic lesions on single‐channel T1w MRI have been made, these approaches have not been systematically evaluated on a large dataset. We performed an exhaustive review of the literature and identified one semiautomated and three fully automated approaches for segmentation of chronic stroke lesions using T1w MRI within the last 10 years: Clusterize, automated lesion identification (ALI), Gaussian naïve Bayes lesion detection (lesionGnb), and lesion identification with neighborhood data analysis (LINDA). We evaluated each method on a large T1w stroke dataset (N = 181). LINDA was the most computationally expensive approach, but performed best across the three main evaluation metrics (median values: dice coefficient = 0.50, Hausdorff's distance = 36.34 mm, and average symmetric surface distance = 4.97 mm). lesionGnb had the highest recall/least false negatives (median = 0.80). However, across the automated methods, many lesions were either misclassified (ALI: 28, lesionGnb: 39, LINDA: 45) or not identified (ALI: 24, LINDA: 23, lesionGnb: 0). Segmentation accuracy in all automated methods were influenced by size (small: worst) and stroke territory (brainstem, cerebellum: worst) of the lesion. To facilitate reproducible science, our analysis files have been made publicly available online.