A comparison study of automated approaches for brain lesions segmentation in ischemic stroke

Stroke lesion segmentation plays a crucial role in neuroimage processing. So far, different methods have been used for brain lesion segmentation. Some of them include a manual tracing method, and others are automated approaches. The majority of these approaches normally require the Magnetic Resonance (MR) imaging sequences as their inputs, and in some cases, tunable parameters should be determined. We use the Ischemic Stroke Lesion Segmentation (ISLES2015) standard MR imaging, which includes training and test MR imaging samples. Here, we consider an ensemble of several approaches to stroke lesion segmentation that are named Lesion Segmentation Toolbox (LST), Automated Lesion Identification (ALI), lesion_Guassian Naïve Bias (GNB), Lesion Identification with Neighborhood Data Analysis (LINDA), DeepMedic, and DeepMedic+Conditional Random Fields (CRF). LST in particular includes two different methods, Lesion Growth Algorithm (LGA) and Lesion Prediction Algorithm (LPA). To examine each approach, the structure of its algorithm is thoroughly discussed; then, the results of these approaches are shown, and the performance of them is evaluated using some statistic parameters. The ultimate results of the parameters are obtained by averaging the results of all samples. Dice similarity coefficient (DC) and Precision of DeepMedic+CRF are more than the other approaches. Regarding the other metrics (Recall, Average Symmetric Surface Distance (ASSD), and Hausdorff's Distance (HD)), the results are a bit different. Among the examined approaches, DeepMedic+CRF yields the best DC than others. Accordingly, the performance of DeepMedic+CRF is better than the rests. Calculations clearly show that the DC of DeepMedic+CRF is approximately 3% higher than single DeepMedic, and on average, 22% more than other approaches.