MC-Net: multi-scale context-attention network for medical CT image segmentation

The encoder-decoder CNN architecture has greatly improved CT medical image segmentation, but it encounters a bottleneck due to the loss of details in the encoding process, which limits the accuracy improvement. To address this problem, we propose a multi-scale context-attention network (MC-Net). The key idea is to explore the useful information across multiple scales and the context for the segmentation of objects of medical interest. Through the introduction of multi-scale and context-attention modules, MC-Net gains the ability to extract local and global semantic information around targets. To further improve the segmentation accuracy, we weight the pixels depending on whether they belong to targets. Many experiments on a lung dataset and a bladder dataset demonstrate that the proposed MC-Net outperforms state-of-the-art methods in terms of accuracy, sensitivity, the area under the receiver operating characteristic curve and the Dice score.