Research on Combining Self-Attention and Convolution for Chest X-Ray Disease Classification

Fig. 5. X-ray annotated images of thoracic disease in the ChestX-ray14 dataset. (a) Atelectasis; (b) cardiomegaly; (c) effusion; (d) infiltration; (e) mass; (f) nodule; (g) pneumonia; (h) pneumothorax

Download full size

Fig. 6. ROC curve and AUC value of 14 diseases on ChestX-ray14 dataset

Download full size

Fig. 7. Results of ablation experiments. (a) Remove ODConv module; (b) remove AC-Block; (c) remove ECA module; (d) remove SAM

Download full size

Table 1. Comparison of different thoracic disease classification algorithms on the ChestX-ray14 dataset

View table

Table 1. Comparison of different thoracic disease classification algorithms on the ChestX-ray14 dataset

Disease	Average AUC
Disease	Ma et al.^［15］	Zhang et al.^［23］	Guan et al.^［14］	Guan et al.^［16］	Chen et al.^［2］	Shao et al.^［24］	Wang et al.^［17］	Ours
Mean	0.794	0.802	0.816	0.822	0.823	0.824	0.826	0.839
Atelectasis	0.763	0.785	0.781	0.785	0.785	0.816	0.779	0.818
Cardiomegaly	0.884	0.876	0.880	0.899	0.887	0.866	0.895	0.909
Effusion	0.816	0.863	0.829	0.835	0.831	0.870	0.836	0.889
Infiltration	0.679	0.673	0.702	0.699	0.703	0.695	0.710	0.716
Mass	0.801	0.804	0.834	0.838	0.833	0.835	0.834	0.844
Nodule	0.729	0.730	0.773	0.775	0.798	0.768	0.777	0.772
Pneumonia	0.710	0.742	0.729	0.738	0.731	0.755	0.737	0.769
Pneumothorax	0.837	0.843	0.857	0.871	0.881	0.869	0.878	0.888
Consolidation	0.744	0.785	0.754	0.763	0.754	0.795	0.759	0.813
Edema	0.841	0.873	0.850	0.850	0.849	0.865	0.855	0.899
Emphysema	0.884	0.858	0.908	0.924	0.930	0.894	0.933	0.928
Fibrosis	0.801	0.775	0.830	0.831	0.833	0.814	0.838	0.825
Pleural thickening	0.754	0.756	0.778	0.776	0.782	0.787	0.791	0.793
Hernia	0.876	0.865	0.917	0.922	0.921	0.903	0.938	0.885

Table 2. Experimental results for module combinations

View table

Table 2. Experimental results for module combinations

Network	Average AUC	Parameter quantity /10⁶	FLOPs /10⁹	Training time /h
Network	Average AUC	Parameter quantity /10⁶	FLOPs /10⁹	Batch size is 16	Batch size is 32
ResNet50	0.764	25.4	3.82	16.10	4.00
ResNet50+ODConv	0.798	27.9	3.99	18.00	5.20
ResNet50+ODConv+ECA	0.810	28.8	4.01	18.60	5.50
ResNet50+ODConv+EDPA	0.815	29.4	4.12	18.70	5.50
ResNet50+ODConv+EDPA+AC-Block	0.839	32.9	6.94	22.00	7.10

Table 3. Experimental results of DA-Net on CheXpert dataset

View table

Table 3. Experimental results of DA-Net on CheXpert dataset

Strategy	Method	Atelectasis	Cardiomegaly	Consolidation	Edema	Pleural effusion	Mean
0	ResNet50	0.806	0.833	0.929	0.913	0.921	0.880
	DenseNet121	0.799	0.832	0.927	0.897	0.923	0.875
	Ours	0.818	0.889	0.941	0.927	0.930	0.901
1	ResNet50	0.805	0.855	0.937	0.910	0.923	0.886
	DenseNet121	0.772	0.844	0.942	0.906	0.901	0.873
	Ours	0.827	0.891	0.948	0.935	0.939	0.908

Tools

Get Citation

Copy Citation Text

Xin Guan, Jingjing Geng, Qiang Li. Research on Combining Self-Attention and Convolution for Chest X-Ray Disease Classification[J]. Laser & Optoelectronics Progress, 2024, 61(4): 0417002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Medical Optics and Biotechnology

Received: Apr. 26, 2023

Accepted: May. 30, 2023

Published Online: Feb. 26, 2024

The Author Email: Qiang Li (liqiang@tju.edu.com)

DOI:10.3788/LOP231180

CSTR:32186.14.LOP231180

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology