Table 2. The model's architecture.

C = Convolutional layer, B = Batch normalization layer, R = Rectified linear unit layer, CN = Concatenation layer, G = Global average pooling layer, D = Dropoutlayer, and F = Fully connected layer.

Layer Number	Filter Size (FS) and Stride (S)	Activations
Input layer	–	224 × 224 × 3
C1, B1, R1	FS = 3 × 3, S = 1	224 × 224 × 32
C2, B2, R2	FS = 5 × 5, S = 2	112 × 112 × 32
C3, B3, R3	FS = 1 × 1, S = 1	112 × 112 × 32
C4, B4, R4	FS = 3 × 3, S = 1	112 × 112 × 32
C5, B5, R5	FS = 5 × 5, S = 1	112 × 112 × 32
C6, B6, R6	FS = 7 × 7, S = 1	112 × 112 × 32
CN1	Five inputs	112 × 112 × 160
B1x	Batch Normalization Layer	112 × 112 × 160
C7, B7, R7	FS = 1 × 1, S = 2	56 × 56 × 64
C8, B8, R8	FS = 3 × 3, S = 2	56 × 56 × 64
C9, B9, R9	FS = 5 × 5, S = 2	56 × 56 × 64
C10, B10, R10	FS = 7 × 7, S = 2	56 × 56 × 64
C11, B11, R11	FS = 3 × 3, S = 2	56 × 56 × 32
CN2	Five inputs	56 × 56 × 228
B2x	Batch Normalization Layer	56 × 56 × 228
G1	–	1 × 1 × 228
F1	200 FC	1 × 1 × 200
D1	Dropout layer with learning rate:0.5	1 × 1 × 200
F2	2 FC	1 × 1 × 2
O (Softmax function)	Normal, Abnormal	1 × 1 × 2