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

_images/voxel.png

Fig 1: Object shape error voxelization


  • Encoder with down-sampling kernels

_images/encoder.png

Fig 1: Encoder with down-sampling kernels


  • Decoder with up-sampling kernels

_images/decoder_interpret.png

Fig 1: Decoder with up-sampling kernels


  • Attention gate

_images/attention.png

Fig 1: Attention gate


  • Residual connections

_images/residuals.png

Fig 1: Residual connections