. 2019 Jul 15;195:285–299. doi: 10.1016/j.neuroimage.2019.01.077

Table 5.

Architecture of the SphericalNet.

#Layer	Type	Parameters	Activation
1	Conversion	From SH to signal space (SH order 4; Laplace-Beltrami Regularization 0.006)	–
2	Spherical Surface Convolution (applied on gradient signals)	Input: 1 Shell; Output: 16 Shells; kernel size: 5; $Θ = \frac{π}{10}$	Sigmoid
3	Spherical Surface Convolution (applied on gradient signals)	Input: 16 Shell; Output: 16 Shells; kernel size: 5; $Θ = \frac{π}{10}$	Sigmoid
4	Spherical Surface Convolution applied on gradient signals)	Input: 16 Shell; Output: 16 Shells; kernel size: 5; $Θ = \frac{π}{10}$	Sigmoid
5	Conversion	From signal to SH space (SH order 4; Laplace-Beltrami Regularization 0.006)
6	Batch Normalization	–	–
7	3D Spatial Convolution	Kernel size: 3 × 3 × 3, padding: 1	PReLU
8	3D Spatial Convolution	Kernel size: 3 × 3 × 3, padding: 1	PReLU
9	3D Spatial Convolution	Kernel size: 3 × 3 × 3, padding: 0