Convolutional neural network -based phantom image scoring for mammography quality control

Sundell, Veli-Matti; Mäkelä, Teemu; Vitikainen, Anne-Mari; Kaasalainen, Touko

doi:10.1186/s12880-022-00944-w

BMC Medical Imaging

Table 1 Convolutional neural network architectures

From: Convolutional neural network -based phantom image scoring for mammography quality control

CNN1 tp: 267,122	CNN2 tp: 386,902	CNN3 tp:236,662	CNN4 tp: 280,082	CNN5 tp: 434,762	CNN6 tp: 748,702	CNN7 tp: 1,025,102	CNN8 tp: 1,366,362
input	input	input	input	input	input	input	input
128 × 128, 5 × 5 conv, 30	128 × 128, 5 × 5 conv, 30	128 × 128, 5 × 5 conv, 30	128 × 128, 5 × 5 conv, 30	128 × 128, 5 × 5 conv, 30	128 × 128, 5 × 5 conv, 30	128 × 128, 5 × 5 conv, 30	128 × 128, 5 × 5 conv, 30
128 × 128 × 30, N + A + P	128 × 128 × 30, N + A + P	128 × 128 × 30, N + A + P	128 × 128 × 30, N + A + P	128 × 128 × 30, N + A + P	128 × 128 × 30, N + A + P	128 × 128 × 30 N + A + P	128 × 128 × 30, N + A + P
63 × 63 × 30, 3 × 3 conv, 45	63 × 63 × 30, 3 × 3 conv, 45	63 × 63 × 30, 3 × 3 conv, 45	63 × 63 × 30, 3 × 3 conv, 45	63 × 63 × 30, 3 × 3 conv, 45	63 × 63 × 30, 3 × 3 conv, 45	63 × 63 × 30, 3 × 3 conv, 45	63 × 63 × 30, 3 × 3 conv, 45
63 × 63 × 45, N + A + P	63 × 63 × 45, N + A + P	63 × 63 × 45, N + A + P	63 × 63 × 45, N + A + P	63 × 63 × 45, N + A + P	63 × 63 × 45, N + A + P	63 × 63 × 45, N + A + P	63 × 63 × 45, N + A + P
31 × 31 × 45, dropout, 30%	31 × 31 × 45, dropout, 30%	31 × 31 × 45, dropout, 30%	31 × 31 × 45, dropout, 30%	31 × 31 × 45, dropout, 30%	31 × 31 × 45, dropout, 30%	31 × 31 × 45, dropout, 30%	31 × 31 × 45, dropout, 30%
31 × 31 × 45, 3 × 3 conv, 60	31 × 31 × 45, 3 × 3 conv, 60	31 × 31 × 45, 3 × 3 conv, 60	31 × 31 × 45, 3 × 3 conv, 60	31 × 31 × 45, 3 × 3 conv, 60	31 × 31 × 45, 3 × 3 conv, 60	31 × 31 × 45, 3 × 3 conv, 60	31 × 31 × 45, 3 × 3 conv, 60
31 × 31 × 60, N + A + P	31 × 31 × 60, N + A + P	31 × 31 × 60, N + A + P	31 × 31 × 60, N + A + P	31 × 31 × 60, N + A + P	31 × 31 × 60, N + A + P	31 × 31 × 60, N + A + P	31 × 31 × 60, N + A + P
15 × 15 × 60, dropout, 50%	15 × 15 × 60, dropout, 30%	15 × 15 × 60, dropout, 30%	15 × 15 × 60, dropout, 30%	15 × 15 × 60, dropout, 30%	15 × 15 × 60, dropout, 30%	15 × 15 × 60, dropout, 30%	15 × 15 × 60, dropout, 30%
15 × 15 × 60, fully connected	15 × 15 × 60, 3 × 3 conv, 80	15 × 15 × 60, 3 × 3 conv, 80	15 × 15 × 60, 3 × 3 conv, 80	15 × 15 × 60, 3 × 3 conv, 80	15 × 15 × 60, 3 × 3 conv, 80	15 × 15 × 60, 3 × 3 conv, 80	15 × 15 × 60, 3 × 3 conv, 80
1 × 1 × 17, softmax	15 × 15 × 80, N + A	15 × 15 × 80, N + A + P	15 × 15 × 80, N + A + P	15 × 15 × 80, N + A + P	15 × 15 × 80, N + A + P	15 × 15 × 80, N + A + P	15 × 15 × 80, N + A + P
	15 × 15 × 80, dropout, 50%	7 × 7 × 80, dropout, 30%	7 × 7 × 80, dropout, 30%	7 × 7 × 80, dropout, 30%	7 × 7 × 80, dropout, 30%	7 × 7 × 80, dropout, 30%	7 × 7 × 80, dropout, 30%
	15 × 15 × 80, fully connected	7 × 7 × 80, 3 × 3 conv, 100	7 × 7 × 80, 3 × 3 conv, 100	7 × 7 × 80, 3 × 3 conv, 100	7 × 7 × 80, 3 × 3 conv, 100	7 × 7 × 80, 3 × 3 conv, 100	7 × 7 × 80, 3 × 3 conv, 100
	1 × 1 × 17, softmax	7 × 7 × 100, N + A	7 × 7 × 100, N + A + P	7 × 7 × 100, N + A + P	7 × 7 × 100, N + A	7 × 7 × 100, N + A	7 × 7 × 100, N + A
		7 × 7 × 100, dropout, 50%	3 × 3 × 100, dropout, 30%	3 × 3 × 100, dropout, 30%	7 × 7 × 100, dropout, 30%	7 × 7 × 100, dropout, 30%	7 × 7 × 100, dropout, 30%
		7 × 7 × 100, fully connected	3 × 3 × 100, 3 × 3 conv, 120	3 × 3 × 100, 3 × 3 conv, 120	7 × 7 × 100, 3 × 3 conv, 120	7 × 7 × 100, 3 × 3 conv, 120	7 × 7 × 100, 3 × 3 conv, 120
		1 × 1 × 17, softmax	3 × 3 × 120, N + A	3 × 3 × 120, N + A	7 × 7 × 120, N + A	7 × 7 × 120, N + A	7 × 7 × 120, N + A
			3 × 3 × 120, dropout, 50%	3 × 3 × 120, dropout, 30%	7 × 7 × 120, dropout, 30%	7 × 7 × 120, dropout, 30%	7 × 7 × 120, dropout, 30%
			3 × 3 × 120, fully connected	3 × 3 × 120, 3 × 3 conv, 140	7 × 7 × 120, 3 × 3 conv, 140	7 × 7 × 120, 3 × 3 conv, 140	7 × 7 × 120, 3 × 3 conv, 140
			1 × 1 × 17 softmax	3 × 3 × 140, N + A	7 × 7 × 140, N + A	7 × 7 × 140, N + A	7 × 7 × 140, N + A
				3 × 3 × 140, dropout, 50%	7 × 7 × 140, dropout, 30%	7 × 7 × 140, dropout, 30%	7 × 7 × 140, dropout, 30%
				3 × 3 × 140, fully connected	7 × 7 × 140, 3 × 3 conv, 160	7 × 7 × 140, 3 × 3 conv, 160	7 × 7 × 140, 3 × 3 conv, 160
				1 × 1 × 17, softmax	7 × 7 × 160, N + A	7 × 7 × 160, N + A	7 × 7 × 160, N + A
					7 × 7 × 160, dropout, 50%	7 × 7 × 160, dropout, 30%	7 × 7 × 160, dropout, 30%
					7 × 7 × 160, fully connected	7 × 7 × 160, 3 × 3 conv, 180	7 × 7 × 160, 3 × 3 conv, 180
					1 × 1 × 17, softmax	7 × 7 × 180, N + A	7 × 7 × 180, N + A
						7 × 7 × 180, dropout, 50%	7 × 7 × 180, dropout, 30%
						7 × 7 × 180, fully connected	7 × 7 × 180, 3 × 3 conv, 200
						1 × 1 × 17, softmax	7 × 7 × 200, N + A
							7 × 7 × 200, dropout, 50%
							7 × 7 × 200, fully connected
							1 × 1 × 17, softmax

Number after convolutional layer (conv) states the number of filters. N + A + P refers to normalization, activation, and pooling block, and N + A refers to normalization and activation block. Input size is given before each layer or layer block. The number of training parameters (tp) is given before network architectures

Back to article page

ISSN: 1471-2342

Contact us

General enquiries: journalsubmissions@springernature.com