Hierarchically Spatial Encoding Module for Chronic Stroke Lesion Segmentation

Since stroke is still the leading cause of death worldwide, neuroimaging research to investigate the mechanism under the brain reorganization and rehabilitation is becoming predominant. In such areas, lesion delineation is a critical step and manual tracing is the gold standard so far, but it is time-consuming and readily influenced by bias. In this work, we propose a volumetric convolutional neural network for lesion segmentation with a novel and well-motivated Hierarchically Spatial Encoding Module (HSEM) based on Recurrent Neural Network (RNN). The proposed HSEM leverages the inherent spatial characteristics of the human brain by encoding the context from various dimensions hierarchically, therefore enhancing the segmentation results. We evaluate our approach on a fresh-new benchmark data set ATLAS. Our method achieves the F1-score of 66.07%, precision of 69.32%, recall of 67.6%, and obtains superior performance compared with other state-of-the-arts. The experiments prove that our proposed module has strong robustness when embedded into the neural network structure and further verify the module can mimic the human brain to make a diagnostic decision with the existence of inherent anisotropy and spatial properties. Furthermore, our model has the potential to provide finer lesion segmentation in some cases where the lesion is not annotated by experts perfectly.