Childhood Pneumonia CT Image Segmentation Network with Prior Graph Convolution and Transformer Fusion

Haocheng Liang; Lü Jia; Mingkai Yu; Xin Chen

doi:10.3788/AOS240772

Acta Optica Sinica, Volume. 44, Issue 16, 1610002(2024)

Childhood Pneumonia CT Image Segmentation Network with Prior Graph Convolution and Transformer Fusion

Haocheng Liang¹, Lü Jia^1,2、*, Mingkai Yu¹, and Xin Chen^3、**

Author Affiliations

¹College of Computer and Information Sciences, Chongqing Normal University, Chongqing 401331, China

²National Center for Applied Mathematics in Chongqing, Chongqing Normal University, Chongqing 401331, China

³Ministry of Education Key Laboratory of Child Development and Disorders, National Clinical Research Center for Child Health and Disorders, Children’s Hospital of Chongqing Medical University, Chongqing 400014, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. Structure of the GTU-Net

Download full size

Fig. 2. Structure of CAM

Download full size

Fig. 3. Process of prior graph learning

Download full size

Fig. 4. Illustration of the ViT block

Download full size

Fig. 5. Childhood pneumonia CT dateset

Download full size

Fig. 6. Ablation results inside PGL module on the Child-P dataset

Download full size

Fig. 7. Visualization results of the prior graph adjacency matrix. (a) Prior graph adjacency matrix without PGL module; (b) prior graph adjacency matrix with PGL module; (c) ideal prior graph

Download full size

Fig. 8. Hyperparametric confusion matrices on the Child-P dataset. (a) JI; (b) SE; (c) MCC; (d) ASD

Download full size

Fig. 9. Segmentation results (up) and their confidence maps (down) on the Child-P, COVID, and MosMed datasets. (a) Label; (b) GTU-Net; (c) U-Net; (d) U-Net++; (e) TMU-Net; (f) TransDeepLab; (g) CSU-Net

Download full size

Fig. 10. Local segmentation results and their heatmaps. (a) Original samples; (b) local labels; (c) GTU-Net; (d) U-Net; (e) TransDeepLab; (f) CSU-Net

Download full size

Fig. 11. Performance comparison of each Transformer network before and after pre-training weight loading on the COVID dataset

Download full size

Fig. 12. Comparison of computational complexity on the Child-P dataset

Download full size

Table 1. Evaluation metrics
View table
Table 1. Evaluation metrics
Metrics Description
DSC /% $2 | G ⋂ \hat{G} | / | G | + | \hat{G} |$
JI /% $| G ⋂ \hat{G} | / | G ⋃ \hat{G} |$
SE /% $P_{T} / (P_{T} + N_{F})$
MCC /% $(P_{T} N_{T} - P_{F} N_{F}) /$ ${[(P_{T} + P_{F}) (P_{T} + N_{F}) (N_{T} + P_{F}) (N_{T} + N_{F})]}^{1 / 2}$
ASD /pixel $\{\sum_{x \in B (G)} d [x, B (\hat{G})] + \sum_{y \in B (\hat{G})} d [y, B (G)]\} /$ $[B (G) + B (\hat{G})]$

Table 2. Ablation results of GTU-Net’s modules on the Child-P dataset

View table

Table 2. Ablation results of GTU-Net’s modules on the Child-P dataset

Baseline: U-Net (32)	CAP	GCT		DSC /%(↑)	JI /%(↑)	SE /%(↑)	MCC /%(↑)	ASD /pixel(↓)
Baseline: U-Net (32)	CAP	PriorGCN	ViT	DSC /%(↑)	JI /%(↑)	SE /%(↑)	MCC /%(↑)	ASD /pixel(↓)
√				0.8368	0.7194	0.7931	0.8356	0.4828
√	√			0.8759	0.7793	0.8497	0.8744	0.4553
√		√		0.8812	0.7876	0.8562	0.8797	0.3213
√			√	0.8751	0.7780	0.8525	0.8734	0.4139
√		√	√	0.8858	0.7951	0.8714	0.8841	0.3508
√	√	√		0.8861	0.7955	0.8607	0.8847	0.2932
√	√		√	0.8803	0.7863	0.8731	0.8784	0.4520
√	√	√	√	0.8923	0.8055	0.8827	0.8906	0.3250

Table 3. Ablation results of PriorGCN replacement (repl.)

View table

Table 3. Ablation results of PriorGCN replacement (repl.)

Module		DSC /%	JI /%	SE /%	MCC /%	ASD /pixel
PriorGCN		0.8923	0.8055	0.8827	0.8906	0.3250
repl. GCN module	GloRe^［17］	0.8840	0.7921	0.8687	0.8823	0.3705
	DualGCN^［19］	0.8843	0.7926	0.8756	0.8825	0.4322
	GCR^［20］	0.8842	0.7925	0.8755	0.8824	0.4196
	SpyGCN^［21］	0.8849	0.7935	0.8740	0.8831	0.4106
repl. Attention module	SE^［29］	0.8696	0.7693	0.8394	0.8681	0.4212
	CBAM^［30］	0.8842	0.7924	0.8800	0.8823	0.4720
	SA^［25］	0.8811	0.7875	0.8635	0.8794	0.3940
	Non-local^［31］	0.8849	0.7936	0.8724	0.8831	0.3810

Table 4. Comparison of metrics on the Child-P dataset of GTU-Net

View table

Table 4. Comparison of metrics on the Child-P dataset of GTU-Net

Network	DSC /%	JI /%	SE /%	MCC /%	ASD /pixel
U-Net^[4]	0.8124	0.6841	0.7394	0.8138	0.3518
ResU-Net^[32]	0.7848	0.6458	0.7138	0.7856	0.6034
U-Net++^[33]	0.8326	0.7133	0.7764	0.8324	0.4418
TMU-Net^[34]	0.7967	0.6621	0.7377	0.7963	0.6594
Swin-Unet^[12]	0.8370	0.7197	0.8208	0.8345	0.6686
SMESwin-Unet^[13]	0.8112	0.6823	0.7909	0.8085	0.8807
TransDeepLab^[14]	0.8574	0.7503	0.8442	0.8552	0.5733
CSU-Net^[35]	0.8741	0.7764	0.8694	0.8721	0.5079
GTU-Net	0.8923	0.8055	0.8827	0.8906	0.3250

Table 5. Comparison of metrics on the COVID and MosMed datasets of GTU-Net

View table

Table 5. Comparison of metrics on the COVID and MosMed datasets of GTU-Net

Network	Metrics on the COVID					Metrics on the MosMed
Network	DSC /%	JI /%	SE /%	MCC /%	ASD /pixel	DSC /%	JI /%	SE /%	MCC /%	ASD /pixel
U-Net	0.7732	0.6303	0.7083	0.7611	0.2804	0.6971	0.5351	0.5842	0.7086	0.3843
ResU-Net	0.7857	0.6471	0.7273	0.7735	0.2660	0.6884	0.5249	0.5748	0.7003	0.3911
U-Net++	0.8004	0.6672	0.7459	0.7887	0.2648	0.7107	0.5512	0.6199	0.7164	0.4837
TMU-Net	0.7892	0.6518	0.7275	0.7776	0.2403	0.6951	0.5327	0.5959	0.7029	0.6031
Swin-Unet	0.7615	0.6149	0.7066	0.7472	0.3839	0.6424	0.4731	0.5585	0.6473	1.2751
SMESwin-Unet	0.6822	0.5177	0.6908	0.6574	0.9243	0.6233	0.4528	0.5489	0.6264	1.3313
TransDeepLab	0.7817	0.6417	0.7269	0.7688	0.3574	0.7273	0.5715	0.6857	0.7265	0.8403
CSU-Net	0.7532	0.6040	0.7138	0.7365	0.4126	0.6712	0.5052	0.5736	0.6790	0.7945
GTU-Net	0.8174	0.6913	0.7862	0.8047	0.2844	0.7415	0.5892	0.6576	0.7458	0.4505

Tools

Get Citation

Copy Citation Text

Haocheng Liang, Lü Jia, Mingkai Yu, Xin Chen. Childhood Pneumonia CT Image Segmentation Network with Prior Graph Convolution and Transformer Fusion[J]. Acta Optica Sinica, 2024, 44(16): 1610002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites