Simplified Automated Segmentation of Acute Ischemic Stroke Lesions from Multimodal MRI: A knowledge-based learning approach

In acute ischemic stroke (AIS), timely detection and reliable segmentation of AIS lesions on magnetic resonance images (MRIs) are essential for patient selection for treatment. Our purpose was to develop and validate an innovative knowledge-based automated technique that utilizes diffusion weighted images (DWIs) and apparent diffusion coefficient (ADC) images for detection and segmentation of AIS lesions. We developed and evaluated our method on a large dataset (n=156) images including 106 stroke and 50 healthy controls. The proposed method compares DWI and ADC images of the subjects - after MNI space normalization and Gaussian kernel smoothing with full width at half maximum (FWHM) of range 2 to 5 mm - with a group of 50 healthy images in voxel-level by means of the Crawford-Howell t-test. T-maps created by the t-test are then screened for extreme values. Areas with hypersignal on DWI and hyposignal on ADC are identified as lesion. We divided our data to train set and validation set and measured Dice Similarity Index (DSI) between automated and manually segmented lesions. The mean DSI for FWHM range from 2 to 5 was 0.47, 0.49, 0.48, and 0.42 on the train set respectively. The method was optimized for thresholds with images smoothed by FWHM of 3, which demonstrated the best DSI. The mean DSI for the optimized parameters on the unseen validation set was 0.50 (SD = 0.19). In conclusion, this fully automated approach showed good agreement with the manually drawn lesions. The approach does not require high computational power, which allows the technique to be implemented on an ordinary computer to be integrated into the routine clinical diagnostic pipeline.