Auto-Segmentation on Liver with U-Net and Pixel De-Convolutional Network

Purpose:？To improve the liver auto-segmentation performance of three-dimensional (3D) U-net by replacing the conventional up-sampling convolution layers with the Pixel De-convolutional Network (PDN) that considers spatial features. Methods: The U-net was originally developed to segment neuronal structure with outstanding performance but suffered serious artifacts from indirectly unrelated adjacent pixels in its up-sampling layers. The hypothesis of this study was that the segmentation quality of the liver could be improved with PDN in which the up-sampling layer was replaced by a pixel de-convolution layer (PDL). Seventy-eight plans of abdominal cancer patients were anonymized and exported. Sixty-two were chosen for training two networks: 1) 3D U-Net, and 2) 3D PDN, by minimizing the Dice loss function. The other sixteen plans were used to test the performance. The similarity Dice and Average Hausdorff Distance (AHD) were calculated and compared between these two networks. Results: The computation time for 62 training cases and 200 training epochs was about 30 minutes for both networks. The segmentation performance was evaluated using the remaining 16 cases. For the Dice score, the mean ± standard deviation were 0.857 ± 0.011 and 0.858 ± 0.015 for the PDN and U-Net, respectively. For the AHD, the mean ± standard deviation were 1.575 ± 0.373 and 1.675 ± 0.769, respectively, corresponding to an improvement of 6.0% and 51.5% of mean and standard deviation for the PDN. Conclusion: The PDN has outperformed the U-Net on liver auto-segmentation. The predicted contours of PDN are more conformal and smoother when compared with the U-Net.