3D U-Net Architecture Global Interpretability¶
Interpretability is achieved using 3D Gradient-Weighted Class Activation Maps (3D Grad-CAMs) for all layers within the architecture. Outputs for various functional elements are shown below.
Object shape error voxelization

Fig 1: Object shape error voxelization¶
Encoder with down-sampling kernels

Fig 1: Encoder with down-sampling kernels¶
Decoder with up-sampling kernels

Fig 1: Decoder with up-sampling kernels¶
Attention gate

Fig 1: Attention gate¶
Residual connections

Fig 1: Residual connections¶