Validating Stroke Lesion Segmentation Methods Using MRI in Children: Transferability of Deep Learning Models

Accurate segmentation of stroke lesions from paediatric brain MRI scans is a challenging task due to the heterogeneity in size, shape and texture of the injuries. Deep learning and atlas-based techniques developed for adult patients may lead to a reduced segmentation performance when applied to children because they do not account for the changes in brain shape that occur during childhood development. The objectives of this work are 2-folds: to investigate the learning transferability of stroke lesion segmentation models trained on the adult domain and applied to the children domain, and to identify the deep learning architecture that produces the most accurate segmentation results.