Table 1.

The dStripe network architecture as used during training. Tensor shape denotes number of channels, and inplane extent. Layers with operations aggregating information across multiple spatial locations are highlighted with their (maximum) spatial extent in brackets, those without are pointwise operations or $1\times 1\times 1$ convolutions. For clarity, tensor shapes are omitted if unchanged. Parameter counts for fixed layers are denoted in brackets. Layers and dynamic range constraint (DRC) and frequency filters are described in sections 2.5.1 and 2.5.2. For inference, layer #15 is deactivated and its function replaced by FFT-based high-pass-filtering of the network output (see Section 2.5.6).

#	layer	tensor shape	param.
	input volume $\mathbf{S}$	1, $n_x$, $n_y$, $n_z$
1	SeparableConv (3,3,3)	16, $n_x$, $n_y$, $n_z$	60
2	BatchNorm		32
3	ConvBlock (3,3,7)		2128
4	ConvBlock (3,3,7) & ReLU		2128
5	ConcatPool (2,2,1)	32, $\frac{n_x}{2}$, $\frac{n_y}{2}$, $n_z$
6	BatchNorm		64
7	ConvBlock (3,3,7)		5792
8	ConvBlock (3,3,7) & ReLU		5792
9	ConcatPool (.,.,1)	64, 16, 16, $n_z$
10	SeparableConv	32, 16, 16, $n_z$	2080
11	SeparableConv & ReLU	1, 16, 16, $n_z$	33
12	DRC: $\mathbf{T}\mapsto \frac{2.0}{1+\exp (-\mathbf{T})}+{10}^{-4}$
13	log transform
14	x,y lowpass filter (9,9,1)		(81)
15	z highpass filter (1,1,9)		(9)
16	exp transform
17	x,y upsample	1, $n_x$, $n_y$, $n_z$