Deep convolutional neural network for automatically segmenting acute ischemic stroke lesion in multi-modality MRI

Abstract

Correct segmentation of stroke lesions from magnetic resonance imaging (MRI) is crucial for neurologists and patients. However, manual segmentation relies on expert experience and is time-consuming. The complicated stroke evolution phase and the limited samples pose challenges for automatic segmentation. In this study, we propose a novel deep convolutional neural network (Res-CNN) to automatically segment acute ischemic stroke lesions from multi-modality MRIs. Our network draws on U-shape structure, and we embed residual unit into network. In Res-CNN, we use residual unit to alleviate the degradation problem and use multi-modality to exploit the complementary information in MRIs. Before training the model, we use data fusion and data augmentation methods to increase the number of training images. Seven neural networks are extensively evaluated on two acute ischemic stroke datasets. Res-CNN shows good performance compared with other six networks both in single modality and multi-modality. Furthermore, compared with the gold standard segmentation manually labeled by two neurologists on a local test dataset, our network achieves the best results in seven neural networks. The average Dice coefficient and Hausdorff distance of our method are 74.20% and 2.33 mm, respectively. Our proposed network may provide a useful tool for segmentation lesion of acute ischemic stroke.