Deep Multimodal Active Contour for Automatic Ischemic Stroke Lesion Segmentation

Ischemic stroke lesion segmentation in Magnetic Resonance Images (MRIs) is a challenging task due to lesion type, shape, and size variations in the brain. Also, MRIs suffer from intensity inhomogeneity and different types of noises. Lesion segmentation is a manual task for clinicians which is error-prone and very time-consuming. This paper aims to present an automatic segmentation for Ischemic stroke lesions in MRIs. The proposed method is a new active contour model based on morphology and fuzzy kernel clustering with a combination of modified U-Net for initial contour selection. In this way, a fully automatic stroke lesion segmentation is obtained. Experiments are conducted on ISLES 2015 dataset. We use T2 and Flair modalities as a vector-valued approach for multimodal segmentation. Results show that our proposed method can segment stroke lesions with proper accuracy. Dice, Jaccard, Sensitivity, and Specificity for automatic lesion segmentation against ground truth are 0.9652 0.0316, 0.9344 0.0561, 0.9903 0.0033 and 0.9903 0.0033 respectively. Our multimodal proposed method shows satisfactory results for Ischemic lesion segmentation due to fuzzy clustering and morphological approach. Moreover, our method is not sensitive to initialization thanks to the proposed modified U-Net. Also, our method is capable to segment images with intensity inhomogeneity.