Sibaji Gaj1, Dennis Chan1, and Xiaojuan Li1
1Department of Biomedical Engineering, Cleveland Clinic, Cleveland, OH, United States
1Department of Biomedical Engineering, Cleveland Clinic, Cleveland, OH, United States
For knee bone and cartilage segmentation, U-Net
deep learning model with softmax activation at last layer was best in terms of ASD.
Adding surface distance to
give more importance of the losses around the boundary regions of the
compartments improved segmentation performance in terms of ASD.
Table 1. Dice coefficients and surface distances for models with
different loss functions and activation functions at different layers. The
highest performances are bold.
Figure 1. Prediction
masks of two models using categorical cross entropy loss with and without
surface distance loss. Adding surface distance loss function significantly
improved the segmentation performance.