An active learning approach for stroke lesion segmentation on multimodal MRI data

The segmentation of lesion tissue in brain images of stroke patients serves to identify the extent of the affected tissues, to perform prognosis on its recovery, and to measure its evolution in longitudinal studies. The different regions of the lesion may have different imaging contrast properties in different image modalities, making difficult the automation of the segmentation process. In this paper we consider an Active Learning selective sampling approach to build image data classifiers from multimodal MRI data to perform voxel based lesion segmentation. We report encouraging results over a dataset combining functional, anatomical and diffusion data.