Improved Attention U-Net Combined with Conditional Random Field for Ischemic Lesion Segmentation from Magnetic Resonance Images

Stroke Lesion segmentation from magnetic resonance images is of great research interest due to its capability in providing appropriate clinical information for the effective treatment of stroke. Deep learning methods have demonstrated promising results in medical image segmentation and U-Net is one of the most effective models. Nevertheless, these algorithms in the area of ischemic stroke lesion segmentation are in the first stages of development and they lack performance compared to other problems for instance brain tumor segmentation. In this paper, we improved the U-Net algorithm by applying blocks consisting of depth-wise separable convolutions with skip connections instead of normal convolution layers and new attention blocks. Based on these improvements, the new architecture has better performance and accuracy with fewer parameters which would need simpler equipment for implementation. We utilized a 3D fully connected Conditional Random Field (CRF) as post-processing to improve the model prediction. Experimental results showed that the proposed end-to-end deep encoder-decoder model has a significant improvement compared to existing deep learning methods on the publicly available Anatomical Tracings of Lesion After Stroke (ATLAS) dataset.