Table 2.

Neural Network Architecture consisting of 17 stages, each of which contains a convolutional layer followed by a number of optional layers including max pooling.

Stage number	Kernel size	Stride	Maxpool	Non-linearity	Batch normalization	Dropout (p = 0.5)	Resulting feature maps	Resulting resolution
Input							35	500 × 250
1	7 × 7	2 × 2		Leaky ReLU	Yes	Spatial	70	250 × 125
2	3 × 3	1 × 1		Leaky ReLU	Yes	No	70	250 × 125
3	3 × 3	1 × 1		Leaky ReLU	Yes	No	70	250 × 125
4	3 × 3	1 × 1	2 × 2	Leaky ReLU	Yes	No	120	125 × 62
5	3 × 3	1 × 1		Leaky ReLU	Yes	No	120	125 × 62
6	3 × 3	1 × 1		Leaky ReLU	Yes	No	120	125 × 62
7	3 × 3	1 × 1	2 × 2	Leaky ReLU	Yes	No	150	62 × 31
8	3 × 3	1 × 1		Leaky ReLU	Yes	No	150	62 × 31
9	3 × 3	1 × 1		Leaky ReLU	Yes	No	150	62 × 31
10	3 × 3	1 × 1	1 × 4	Leaky ReLU	Yes	No	150	62 × 7
11	3 × 3	1 × 1		Leaky ReLU	Yes	No	150	62 × 7
12	3 × 3	1 × 1		Leaky ReLU	Yes	No	150	62 × 7
13	3 × 3	1 × 1	1 × 4	Leaky ReLU	Yes	No	150	62 × 1
14	1 × 1	1 × 1		Leaky ReLU	No	Yes	100	62 × 1
15	1 × 1	1 × 1		Leaky ReLU	No	Yes	50	62 × 1
16	1 × 1	1 × 1		Leaky ReLU	No	Yes	20	62 × 1
17	1 × 1	1 × 1		Sigmoid	No	No	1	62 × 1

SpatialDropout was employed in stage 1.