Magnetic Resonance Brain Tumor Image Segmentation Based on Attention U-Net

Lingmei Ai; Tiandong Li; Fuyuan Liao; Kangzhen Shi

doi:10.3788/LOP57.141030

Journals >Laser & Optoelectronics Progress >Volume 57 >Issue 14 >Page 141030 > Article

Laser & Optoelectronics Progress
Vol. 57, Issue 14, 141030 (2020)

Magnetic Resonance Brain Tumor Image Segmentation Based on Attention U-Net

Lingmei Ai^1、*, Tiandong Li^1、**, Fuyuan Liao², and Kangzhen Shi¹

Author Affiliations

¹School of Computer Science, Shaanxi Normal University, Xi'an, Shaanxi 716000, China

²School of Electronic Information Engineering, Xi'an Technological University, Xi'an, Shaanxi 716000, China

show less

DOI: 10.3788/LOP57.141030 Cite this Article Set citation alerts

Lingmei Ai, Tiandong Li, Fuyuan Liao, Kangzhen Shi. Magnetic Resonance Brain Tumor Image Segmentation Based on Attention U-Net[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141030 Copy Citation Text

show less

Fig. 1. Network structure. (a) FCN; (b) U-Net; (c) Res-U-Net; (d) proposed method

Download full size

Fig. 2. Convolution module of the network structure. (a) Residure modules; (b) dense module; (c) residure-dense module

Download full size

Fig. 3. Aattention module of SE-Net. (a) SE-Net attention unit; (b) proposed attention unit

Download full size

Fig. 4. Experimental data. (a) TI image; (b) T1ce image; (c) T2 image; (d) FLAIR image; (E) ground truth

Download full size

Fig. 5. Experimental results of the four models at three different levels of LGG and HGG. (a) LGG; (b) HGG

Download full size

Fig. 6. Comparison of network loss changes under different epoch weighting factors

Download full size

Model	Dice score			Sensitivity			Specificity
Model	WT	TC	ET	WT	TC	ET	WT	TC	ET
FCN	83.16	73.34	63.13	88.75	77.25	71.36	99.83	99.82	99.95
U-Net	84.16	75.23	65.13	89.15	78.15	73.26	99.74	99.80	99.94
Res-U-Net	87.10	74.97	78.32	89.85	79.32	77.42	99.96	99.93	99.97
Proposed	90.56	79.82	78.61	90.25	80.29	78.95	99.98	99.95	99.98

Table 1. Results of the four models%

Model	Dice score			Sensitivity			Specificity
Model	WT	TC	ET	WT	TC	ET	WT	TC	ET
U-Net	84.16	75.23	65.13	89.15	78.15	73.26	99.74	99.80	99.94
RDB+U-Net	87.25	76.20	77.23	87.53	79.42	77.02	99.85	98.90	99.13
TRB+RDB+U-Net	87.38	75.97	77.42	89.52	78.58	77.45	98.86	98.92	99.21
A1+TRB+RDB+U-Net	89.05	78.53	78.46	89.97	79.92	78.51	99.32	99.91	99.42
A2+TRB+RDB+U-Net	88.82	77.25	78.35	89.51	79.36	78.01	99.06	99.88	99.35
A3+TRB+RDB+U-Net	88.23	76.82	77.32	88.97	78.86	77.81	99.85	99.81	99.29
Proposed	90.56	79.82	78.61	90.25	80.29	78.95	99.98	99.95	99.98

Table 2. Results of adding different blocks to the U-NET structure%

Method	Dice_WT	Dice_ET	Dice_TC
Dong et al.^[7]	0.8600	0.6500	0.8600
Isensee et al.^[21]	0.8580	0.6470	0.7750
Puch et al.^[22]	0.8970	0.7520	0.7970
Proposed	0.9056	0.7861	0.7982

Table 3. Dice comparison between proposed method and other advanced segmentation methods

Lingmei Ai, Tiandong Li, Fuyuan Liao, Kangzhen Shi. Magnetic Resonance Brain Tumor Image Segmentation Based on Attention U-Net[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141030

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information