Table 2.

Detailed description of architecture of the stacked CNN-RNN network in our study.

Repeat Times	Layer Type	Padding Size	Stride	Filter Size	Number of Filters (Neurons)	Size of Feature Maps	Number of Parameters
1	Input Layer	n/a	n/a	n/a	n/a	5 $\times$ 224 $\times$ 224 $\times$ 3	0
2	Convolution	1 $\times$ 1	1 $\times$ 1	3 $\times$ 3	64	5 $\times$ 224 $\times$ 224 $\times$ 64	38,720
	ReLU	n/a	n/a	n/a	n/a	5 $\times$ 224 $\times$ 224 $\times$ 64	0
1	Max Pooling	n/a	2 $\times$ 2	2 $\times$ 2	1	5 $\times$ 112 $\times$ 112 $\times$ 64	0
2	Convolution	1 $\times$ 1	1 $\times$ 1	3 $\times$ 3	128	5 $\times$ 112 $\times$ 112 $\times$ 128	221,440
	ReLU	n/a	n/a	n/a	n/a	5 $\times$ 112 $\times$ 112 $\times$ 128	0
1	Max Pooling	n/a	2 $\times$ 2	2 $\times$ 2	1	5 $\times 56\times 56\times$ 128	0
4	Convolution	1 $\times$ 1	1 $\times$ 1	3 $\times$ 3	256	5 $\times 56\times 56\times$ 256	2,065,408
	ReLU	n/a	n/a	n/a	n/a	5 $\times$ 56 $\times$ 56 $\times$ 256	0
1	Max Pooling	n/a	2 $\times$ 2	2 $\times$ 2	1	5 $\times$ 28 $\times$ 28 $\times 256$	0
4	Convolution	1 $\times$ 1	1 $\times$ 1	3 $\times$ 3	512	5 $\times$ 28 $\times$ 28 $\times$ 512	8,259,584
	ReLU	n/a	n/a	n/a	n/a	5 $\times$ 28 $\times$ 28 $\times$ 512	0
1	Max Pooling	n/a	2 $\times$ 2	2 $\times$ 2	1	5 $\times$ 14 $\times$ 14 $\times$ 512	0
4	Convolution	1 $\times$ 1	1 $\times$ 1	3 $\times$ 3	512	5 $\times$ 14 $\times$ 14 $\times$ 512	9,439,232
	ReLU	n/a	n/a	n/a	n/a	5 $\times$ 14 $\times$ 14 $\times$ 512	0
1	Max Pooling	n/a	2 $\times$ 2	2 $\times$ 2	1	5 $\times$ 7 $\times$ 7 $\times$ 512	0
1	Global Average Pooling	n/a	n/a	n/a	1	5 $\times$ 512	0
1	Fully Connected Layer	n/a	n/a	n/a	1024	5 $\times$ 1024	525,312
1	Batch Normalization	n/a	n/a	n/a	n/a	5 $\times$ 1024	4096
1	ReLU	n/a	n/a	n/a	n/a	5 $\times$ 1024	0
1	LSTM	n/a	n/a	n/a	n/a	1024	8,392,704
1	Dropout	n/a	n/a	n/a	n/a	1024	0
1	Fully Connected Layer	n/a	n/a	n/a	2	2	2050
Total number of parameters: 28,948,546 Total number of trainable parameters: 28,946,498 Total number of non-trainable parameters: 2048