Voxel-based Gaussian naive Bayes classification of ischemic stroke lesions in individual T1-weighted MRI scans

Background: Manual lesion delineation by an expert is the standard for lesion identification in MRI scans, but it is time-consuming and can introduce subjective bias. Alternative methods often require multi-modal MRI data, user interaction, scans from a control population, and/or arbitrary statistical thresholding. New method: We present an approach for automatically identifying stroke lesions in individual T1-weighted MRI scans using nave Bayes classification. Probabilistic tissue segmentation and image algebra were used to create feature maps encoding information about missing and abnormal tissue. Leave-one-case-out training and cross-validation was used to obtain out-of-sample predictions for each of 30 cases with left hemisphere stroke lesions. Results: Our method correctly predicted lesion locations for 30/30 un-trained cases. Post-processing with smoothing (8 mm FWHM) and cluster-extent thresholding (100 voxels) was found to improve performance. Comparison with existing method: Quantitative evaluations of post-processed out-of-sample predictions on 30 cases revealed high spatial overlap (mean Dice similarity coefficient = 0.66) and volume agreement (mean percent volume difference = 28.91; Pearson's r = 0.97) with manual lesion delineations. Conclusions: Our automated approach agrees with manual tracing. It provides an alternative to automated methods that require multi-modal MRI data, additional control scans, or user interaction to achieve optimal performance. Our fully trained classifier has applications in neuroimaging and clinical contexts. (C) 2015 Elsevier B.V. All rights reserved.