Lightweight Brain Tumor Segmentation Algorithm Based on Multi-View Convolution

Fig. 2. Schematic of MVC unit. (a) MVC unit when feature numbers of the input equals to the output; (b) MVC unit when feature numbers of the input does not equal to the output; (c) multiplexer module

Download full size

Fig. 3. Schematic of pseudo three-dimensional convolution kernel convolution from different views. (a) Axial; (b) sagittal; (c) coronal

Download full size

Fig. 4. Comparison of segmentation results. (a) T1; (b) T1c; (c) T2; (d) FLAIR; (e) segmentation results of the proposed model; (f) expert manual segmentation results

Download full size

Fig. 5. Schematic of weight parameters changing with training times

Download full size

Table 1. Parameter setting during model training
View table
Table 1. Parameter setting during model training
Parameter Content
Weight decay coefficient 0.00001
Weight initialization He initialization^［23］
Initial learning rate 0.001
Optimizer Adam
Training times 600
Batch size 12

Table 2. Segmentation effect comparison of various algorithms on BraTS validation set in 2019

View table

Table 2. Segmentation effect comparison of various algorithms on BraTS validation set in 2019

Algorithm	Parameters /10⁶	FLOPs /10⁹	Dice /%			HD95 /mm
Algorithm	Parameters /10⁶	FLOPs /10⁹	ET	WT	TC	ET	WT	TC
Proposed algorithm	0.71	6.64	78.32	89.63	82.1	3.22	4.78	6.20
Algorithm in Ref.［11］	4.68	199.69	80.21	90.94	86.47	3.15	4.26	5.44
Algorithm in Ref.［12］			75.40	91.00	83.50	3.84	4.57	5.58
Algorithm in Ref.［13］			77.00	91.00	83.00	3.92	4.52	6.27
Algorithm in Ref.［14］	5.90	1534.99	75.57	90.29	79.32	4.77	4.49	8.19
Algorithm in Ref.［15］	33.55	293.76	72.31	88.82	78.33	4.91	8.12	7.56
Algorithm in Ref.［16］			75.90	89.30	80.70	4.19	6.94	7.66
Algorithm in Ref.［17］			66.68	85.27	70.91	7.27	8.08	9.57
Algorithm in Ref.［24］	3.88	27.04	77.60	90.00	81.50	2.99	4.64	6.22
Algorithm in Ref.［18］	13.08	233.36	77.60	88.40	79.60	4.48	9.11	8.68

Table 3. Experimental environment and settings of various algorithms

View table

Table 3. Experimental environment and settings of various algorithms

Algorithm	Experiment device	Batch size	Patch size
Proposed algorithm	Three parallel Nvidia GTX2080Ti（11 GB）GPUs	12	$128 \times 128 \times 128$
Algorithm in Ref.［11］	One Nvidia Titan V GPU with 12 GB	1	$128 \times 128 \times 128$
Algorithm in Ref.［12］	Two Titan GPUs with 12 GB	1	$128 \times 128 \times 128$
Algorithm in Ref.［13］	One Nvidia Tesla V100 32 GB GPU	1	160×224×160
Algorithm in Ref.［15］	One Nvidia Titan Xp 12 GB GPU	1	$128 \times 128 \times 128$
Algorithm in Ref.［16］	Two GeForce GTX 1080 Ti（11 GB）	2	$128 \times 128 \times 128$
Algorithm in Ref.［24］	Four parallel Nvidia GTX2080Ti（11 GB）GPUs	12	$128 \times 128 \times 128$
Algorithm in Ref.［18］	Three parallel Nvidia GTX2080Ti（11 GB）GPUs	9	$128 \times 128 \times 128$

Table 4. Comparison of segmentation result of different MVC units

View table

Table 4. Comparison of segmentation result of different MVC units

Parameter	Area	Proposed algorithm	Model 1	Model 2	Model 3	Model 4	Model 5
Dice /%	WT	89.63	89.10	89.47	89.54	90.08	89.71
	TC	82.10	80.00	80.78	81.97	80.18	81.29
	ET	78.32	76.19	76.50	77.02	76.30	77.05
HD95 /mm	WT	4.78	7.61	7.70	5.30	4.52	5.09
	TC	6.20	9.00	15.89	5.76	6.39	6.68
	ET	3.22	32.41	35.40	2.95	3.40	2.84

Tools

Get Citation

Copy Citation Text

Chengxiang Shan, Qiang Li, Xin Guan. Lightweight Brain Tumor Segmentation Algorithm Based on Multi-View Convolution[J]. Laser & Optoelectronics Progress, 2023, 60(10): 1010018

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites