Disease Classification Algorithm of Chest X-Ray Based on Efficient Channel Attention

Lingyun Shao; Qiang Li; Xin Guan; Xuewen Ding

doi:10.3788/LOP220759

Journals >Laser & Optoelectronics Progress >Volume 60 >Issue 12 >Page 1217001 > Article

Laser & Optoelectronics Progress
Vol. 60, Issue 12, 1217001 (2023)

Disease Classification Algorithm of Chest X-Ray Based on Efficient Channel Attention

Lingyun Shao¹, Qiang Li¹, Xin Guan^1,*, and Xuewen Ding²

Author Affiliations

¹School of Microelectronics, Tianjin University, Tianjin 300072, China

²Tianjin Fieldbus Control Technology Engineering Center, Tianjin Vocational and Technical Normal University, Tianjin 300222, China

show less

DOI: 10.3788/LOP220759 Cite this Article Set citation alerts

Lingyun Shao, Qiang Li, Xin Guan, Xuewen Ding. Disease Classification Algorithm of Chest X-Ray Based on Efficient Channel Attention[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1217001 Copy Citation Text

show less

Fig. 1. Framework diagram of chest disease classification network

Download full size

Fig. 2. Structure of DECA

Download full size

Fig. 3. 5-layer dense connected block, each layer taking all the preceding feature-maps as input

Download full size

Fig. 4. Efficient channel attention module

Download full size

Fig. 5. Schematic diagram of asymmetric convolution

Download full size

Fig. 6. X-ray images in Chest X-ray 15 dataset. (a) No finding; (b) pneumonia;(c) COVID-19; (d) cardiomegaly; (e) hernia; (f) infiltration;(g) nodule; (h) emphysema; (i) effusion; (j) pleural thickening; (k) pneumothorax; (l) mass; (m) fibrosis; (n) edema; (o) consolidation

Download full size

Fig. 7. ROC curve and AUC value of proposed algorithm. (a) Atelectasis; (b) cardiomegaly; (c) effusion; (d) infiltration; (e) mass; (f) nodule; (g) pneumonia; (h) pneumothorax; (i) consolidation; (j) edema; (k) emphysema; (l) fibrosis; (m) pleural thickening: (n) hernia; (o) COVID-19

Download full size

Layer	Output size	DECANet-121
Convolution	112×112	7×7 Conv，stride 2
Pooling	56×56	3×3 max pool，stride 2
DECA block 1	56×56	$\{\begin{array}{l} G A P \\ 1 \times 1 C o n v \times k \\ S i g m o i d \\ 1 \times 1 C o n v \\ 3 \times 3 A C B \end{array}\} \times 6$
Transition layer 1	56×56	1×1 Conv
Transition layer 1	28×28	2×2 average pool，stride 2
DECA block 2	28×28	$\{\begin{array}{l} G A P \\ 1 \times 1 C o n v \times k \\ S i g m o i d \\ 1 \times 1 C o n v \\ 3 \times 3 A C B \end{array}\} \times 12$
Transition layer 2	28×28	1×1 Conv
Transition layer 2	14×14	2×2 average pool，stride 2
DECA block 3	14×14	$\{\begin{array}{l} G A P \\ 1 \times 1 C o n v \times k \\ S i g m o i d \\ 1 \times 1 C o n v \\ 3 \times 3 A C B \end{array}\} \times 24$
Transition layer 3	14×14	1×1 Conv
Transition layer 3	7×7	2×2 average pool，stride 2
DECA block 4	7×7	$\{\begin{array}{l} G A P \\ 1 \times 1 C o n v \times k \\ S i g m o i d \\ 1 \times 1 C o n v \\ 3 \times 3 A C B \end{array}\} \times 16$
Classification layer	1×1	7×7 global average pool，stride 2
Classification layer		15 fully-connected

Table 1. Specific structure of DECA-Net

Network	Backbone	Average AUC
ResNet50	ResNet50	0.7468
ResNet50+SE	ResNet50	0.7642
ResNet50+ECA DenseNet121	ResNet50 DenseNet121	0.7886 0.7952
DenseNet121+SE^［10］	DenseNet121	0.8014
DECA-Net	DenseNet121	0.8245

Table 2. Comparison of classification results of different classification network models on Chest X-ray 15

Disease	Algorithm of reference［9］	Algorithm of reference［10］	Algorithm of reference［13］	Algorithm of reference［18］	Algorithm of reference［19］	Algorithm of reference［20］	Algorithm of reference［21］	Proposedalgorithm
Atelectasis	0.7003	0.7627	0.785	0.767	0.783	0.791	0.785	0.8157
Cardiomegaly	0.81	0.8835	0.8766	0.883	0.884	0.898	0.887	0.8657
Effusion	0.7585	0.8159	0.8628	0.828	0.832	0.873	0.831	0.8701
Infiltration	0.6614	0.6786	0.673	0.709	0.708	0.700	0.703	0.6948
Mass	0.6933	0.8012	0.804	0.821	0.837	0.832	0.833	0.8350
Nodule	0.6687	0.7293	0.7299	0.758	0.800	0.758	0.798	0.7683
Pneumonia	0.658	0.7097	0.7423	0.731	0.735	0.767	0.731	0.7548
Pneumothorax	0.7993	0.8377	0.8426	0.846	0.866	0.859	0.881	0.8687
Consolidation	0.7032	0.7443	0.7846	0.745	0.746	0.800	0.754	0.7952
Edema	0.8052	0.8414	0.8727	0.835	0.841	0.889	0.849	0.8647
Emphysema	0.833	0.8836	0.858	0.895	0.937	0.891	0.930	0.8942
Fibrosis	0.7859	0.8007	0.7754	0.818	0.82	0.789	0.833	0.8141
Pleural thickening	0.6835	0.7536	0.7563	0.761	0.796	0.771	0.782	0.7872
Hernia	0.8717	0.8763	0.8645	0.896	0.895	0.896	0.921	0.9027
Mean	0.7451	0.7941	0.802	0.807	0.82	0.822	0.823	0.8237

Table 3. Comparison of AUC value of different chest disease classification algorithms on Chest X-ray 14 dataset

Disease	Network_1	Network_2	Network_3	Network_4	DECA-Net
Atelectasis	0.7853	0.7838	0.7974	0.8073	0.8101
Cardiomegaly	0.8770	0.8702	0.8813	0.8877	0.8919
Effusion	0.8541	0.8498	0.8614	0.8689	0.8701
Infiltration	0.6696	0.6724	0.6845	0.6946	0.6892
Mass	0.8128	0.8090	0.8306	0.8302	0.8301
Nodule	0.739	0.7244	0.7595	0.7534	0.7672
Pneumonia	0.7282	0.7355	0.7450	0.7412	0.7389
Pneumothorax	0.8546	0.8623	0.8743	0.8806	0.8829
Consolidation	0.7818	0.7849	0.7908	0.7975	0.7931
Edema	0.8633	0.8872	0.8726	0.8805	0.8782
Emphysema	0.8742	0.8829	0.8924	0.9023	0.9122
Fibrosis	0.7722	0.7892	0.7859	0.796	0.8054
Pleural thickening	0.7607	0.7788	0.7627	0.7707	0.7851
Hernia	0.8571	0.9011	0.9004	0.8678	0.8893
COVID-19	0.8265	0.8157	0.8273	0.8357	0.8239
Mean	0. 8038	0.8098	0.8177	0.8210	0.8245

Table 4. Comparison of ablation experiment results

Lingyun Shao, Qiang Li, Xin Guan, Xuewen Ding. Disease Classification Algorithm of Chest X-Ray Based on Efficient Channel Attention[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1217001

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享