Bottleneck feature supervised U-Net for pixel-wise liver and tumor segmentation

Liver cancer is one of the most common cancer types with high death rate. Doctors diagnose cancer by examining the CT images, which can be time-consuming and prone to error. Therefore, an automatic segmentation method is desired for clinical practice. In the literature, many U-Net-based models were proposed. But few of them focus on the bottleneck feature vectors, which are low dimensional representations of the input. In this paper, we propose a bottleneck feature supervised (BS) U-Net model and apply it to liver and tumor segmentation. Our main contributions are: (1) we propose a variation of the original U-Net that has better performance with a smaller number of parameters; (2) we propose a bottleneck feature supervised (BS) U-Net that contains an encoding U-Net and a segmentation U-Net. The encoding U-Net is first trained as an auto-encoder to get encodings of the label maps, which are subsequently used as additional supervision to train the segmentation U-Net. Compared with most U-Net-based models in the literature that only use the pair information between images and label maps, BS U-Net additionally uses the information extracted from the label maps as supervision. The model is evaluated on the liver and tumor segmentation (LiTS) competition. 2D BS U-Net achieves dice per case (DPC) 96.1% for liver segmentation and 56.9% for tumor segmentation. This result is better than most state-of-the-art 2D UNet-based networks in both tasks. Furthermore, the idea of bottleneck feature supervision can also be generalized to other U-Net-based models, making it have good potential for future development.