Saliency Transformation Network: Incorporating Multi-stage Visual Cues for Pancreas Segmentation

We aim at segmenting small organs (e.g., the pancreas) from abdominal CT scans. As the target often occupies a relatively small region in the input image, deep neural networks can be easily confused by the complex and variable background. To alleviate this, researchers proposed a coarse-to-fine approach (Zhou et al. 2016), which used prediction from the coarse stage to shrink the input region provided to the fine stage. Although this strategy achieves high accuracy, we note that the coarse-scaled and fine-scaled networks were trained and tested individually, which limited the use of multi-stage visual cues for segmentation. This paper presents a Saliency Transformation Network, which contains a trainable saliency transformation module. This module computes spatial weights from the coarse-scaled segmentation score map, and applies them to the fine-scaled input image. In training, the coarse-scaled and fine-scaled segmentation networks are optimized in a joint manner, so that both networks become more powerful when they are evaluated individually. In testing, this strategy makes full use of the segmentation results at the coarse stage, so that we can deliver complementary information to the fine stage rather than merely providing a bounding box. We perform experiments on the NIH pancreas segmentation dataset with 82 CT volumes. Following the same testing process which involves a coarse-to-fine iteration, our approach outperforms the state-of-the-art approach (trained in a stage-wise manner) by an average of over 2%. In addition, our approach enjoys better convergence properties, making it more reliable in practice.