Classifiers for Ischemic Stroke Lesion Segmentation: A Comparison Study.

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Ischemic stroke causes brain tissue infarction due to blocked cerebral blood supply.
• Accurate lesion volume segmentation is crucial for clinical trials but challenging due to lesion variability.
• Automatic segmentation of ischemic stroke lesions is an area of significant research interest.

Purpose of the Study:

• To evaluate and compare the accuracy and reliability of nine classification methods for ischemic stroke lesion segmentation.
• To assess multi-spectral versus mono-spectral classification approaches using multiparametric MRI data.
• To benchmark automated methods against expert human segmentation for inter-observer agreement.

Main Methods:

• Utilized 37 multiparametric MRI datasets from sub-acute ischemic stroke patients.
• Compared nine classification algorithms including Generalized Linear Models, Random Decision Forests, and Convolutional Neural Networks.
• Evaluated both multi-spectral and mono-spectral (FLAIR MRI) classification, with expert segmentations for reference.

Main Results:

• High-level machine learning methods significantly outperformed simpler classification techniques.
• Random Decision Forests and Convolutional Neural Networks achieved the best segmentation results, surpassing previous benchmarks.
• No automated method reached the level of agreement observed between human observers.

Conclusions:

• Complex, non-linear machine learning approaches are superior for ischemic stroke lesion segmentation.
• While advanced methods show improved performance, automatic segmentation remains a challenging problem.
• Further research is necessary to enhance the accuracy and reliability of automated lesion segmentation methods.