Room: Stars at Night Ballroom 2-3
Purpose: This abstract proposes a modified U-Net with incorporation of high-resolution high-level features for liver-tumor segmentation.
Methods: The conventional U-Net has a skip connection to transfer high-resolution low-level features in the encoding part to the decoding part. So, the network can overcome resolution loss by the pooling operation. However, high-level features extracted by the network often do not contain enough high-resolution information of the input, leading to greater uncertainty such as liver-tumor segmentation. Furthermore, the skip connection also occurs duplication of low-resolution information, which leads to blurring extracted image features, as reported by previous studies. To cope with these problems, a residual path is directly combined with the skip connection, and the skip connection has its own convolutional layer. To match the matrix size of features, the residual path has interpolation operation as well. After the residual path, the only edge-like features can be transferred across the skip connection. So, low-resolution features do not pass across the skip connection anymore, and duplication of low-resolution information can be avoided. In addition, a convolutional layer in the skip connection generates higher-lever features of edge-like features passing across the skip connection. So, high-level features can contain more high-resolution information. The validation was performed with the public LiTS-challenge-2017 dataset.
Results: The proposed method offers more precise segmentation results as compared to the conventional U-Net without critical computation complex. The proposed network is working well even though the boundary shape of tumor is irregular and fuzzy. The dice similarity coefficients (DSCs) of the conventional U-Net and the proposed network are 54.84 % and 87.01 %, respectively, and ranked 1st as of Feb. 2019.
Conclusion: The proposed network can be an important steps in radiation therapy of hepatocellular carcinoma. The proposed network can also be improved with real patient data by a kind of reinforce learning.