Real-time semantic segmentation of microvascular decompression images based on encoder-decoder structure

Layer	Operator	Mode	Channel	Output size
1	$3 \times 3$ Conv	stride 2	32	$256 \times 256$
2	$3 \times 3$ Conv	stride 1	32	$256 \times 256$
3	$3 \times 3$ Conv	stride 1	32	$256 \times 256$
4-5	$n \times $LABM	dilated 2	32	$256 \times 256$
6	$3 \times 3$ Conv	stride 2	64	$128 \times 128$
7-8	$m \times $LABM	dilated 4	64	$128 \times 128$
9	$3 \times 3$ Conv	stride 2	128	$64 \times 64$
10-12	$l \times $LABM	dilated 8	128	$64 \times 64$
13	1×FFM	−	128	$64 \times 64$
14	1×FFM	−	64	$128 \times 128$
15	1×FFM	−	32	$256 \times 256$
16	1×1 Conv	stride 1	10	$256 \times 256$
17	Bilinear interpolation	$ \times 2$	10	$512 \times 512$

Table 2. Abbreviations of medical terms and corresponding color

View table

View in Article

Table 2. Abbreviations of medical terms and corresponding color

简称	全称	对应颜色
cn5	三叉神经
cn7	面神经
cn9	舌咽神经
cn10	迷走神经
aica+cn7	小脑前下动脉及面神经
pica+cn7	小脑后下动脉及面神经
pica	小脑后下动脉
aica	小脑前下动脉
pv	岩静脉

Table 3. Training parameters
View table
View in Article
Table 3. Training parameters
Parameter name Parameter selection
Learning rate Policy Initialization Power
poly 0.16 0.9
Optimizer Policy Momentum Weight decay
SGD 0.9 $1\times10 ^{- 4}$
Enter picture size $768 \times 576$
Batch size 8

Table 4. Results of LABM encoder with different combinations of dilation rates
View table
View in Article
Table 4. Results of LABM encoder with different combinations of dilation rates
Name Dilation rates mIoU(%)
LABM_N2M2L4 2,4,8 72.35
LABM_N2M2L4 4,8,16 72.08

Table 5. Results of LABM encoder with different settings
View table
View in Article
Table 5. Results of LABM encoder with different settings
Concatenation Params(M) FLOPs(G) mIoU(%)
0.30 2.81 72.35
√ 0.54 4.03 73.08

Table 6. Results of encoder with different depths when the input size is 512 × 512
View table
View in Article
Table 6. Results of encoder with different depths when the input size is 512 × 512
n m l Params(M) FLOPs(G) mIoU(%)
2 2 2 0.52 3.95 72.35
2 2 4 0.54 4.03 73.08
2 4 4 0.55 4.11 73.84
4 4 4 0.55 4.20 73.37

Table 7. Results of FFM decoder with different components
View table
View in Article
Table 7. Results of FFM decoder with different components
FFM Pooling mIoU(%)
w/o − 73.84
w 77.11
w √ 77.34

Table 8. Effect of dilation of U-MVDNet on mIoU
View table
View in Article
Table 8. Effect of dilation of U-MVDNet on mIoU
Concatenation mIoU(%) Params(M)
U-MVDNet 77.34 0.66
U-MVDNet_w/o dilation 75.61 0.66
U-MVDNet_First $3 \times 3$ conv ( $r = 2$) 76.81 0.66

Table 9. Experimental results on MVD test set

View table

View in Article

Table 9. Experimental results on MVD test set

Method	Params(M)	Speed(frame·s⁻¹)	mIoU(%)	cn5	cn7	cn9	cn10	aica+cn7	pica+cn7	pica	aica	pv
CGNet^[28]	0.94	87.4	71.95	81.26	82.9	71.29	69.85	71.64	87.16	67.37	65.66	50.42
EDANet^[29]	0.69	125	74.51	83.03	84.02	70.31	77.25	75.09	87.98	70.37	68.18	54.34
ContextNet^[30]	0.88	163.3	75.81	82.14	84.15	74.91	78.08	76.67	87.84	72.08	69.77	56.65
U-MVDNet	0.66	140.8	76.29	82.25	85.45	74.8	76.91	76.32	87.85	74.08	69.83	59.12

Table 10. Experimental results on ISIC 2016 + PH2 test set

View table

View in Article

Table 10. Experimental results on ISIC 2016 + PH2 test set

Model	Params (M)	Speed (frame·s⁻¹)	DIC (%)	JAC (%)	ACC (%)	SPE (%)	SEN (%)
DeepLabv3^[31]	58.2	98.7	88.6	81.2	91.9	89.1	95.9
CA-Net^[32]	2.79	130.3	88.7	80.5	93.2	91.3	96.9
U-MVDNet	0.66	175.1	89.3	81.7	93.2	93.3	94.3

Table 11. Two different hardware environments
View table
View in Article
Table 11. Two different hardware environments
Jetson Xavier 服务器
GPU Volta GTX 2080Ti
CPU 8核Carmel ARM 8核i7-9700K
显存 32GB LPDDR4x 11GB GDDR6
显存带宽 136.5 GB/s 616 GB/s
CUDA核心 512 4352

Table 12. Test results by different methods with different resolutions

View table

View in Article

Table 12. Test results by different methods with different resolutions

Method	Size	Times/ms	Speed/frame·s⁻¹	mIoU/%
CGNet[28]	$640 \times 480$	65.7	15.2	70.31
CGNet[28]	$768 \times 576$	69.2	14.4	71.95
EDANet[29]	$640 \times 480$	42.3	23.6	73.2
EDANet[29]	$768 \times 576$	45.2	22.1	74.18
ContextNet[30]	$640 \times 480$	34.5	28.9	74.81
ContextNet[30]	$768 \times 576$	36.1	27.7	75.81
U-MVDNet	$640 \times 480$	41.5	24.2	75.76
U-MVDNet	$768 \times 576$	43.6	22.9	76.29

Tools

Get Citation

Copy Citation Text

Rui-feng BAI, Shan JIANG, Hai-jiang SUN, Xin-rui LIU. Real-time semantic segmentation of microvascular decompression images based on encoder-decoder structure[J]. Chinese Optics, 2022, 15(5): 1055

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Original Article

Received: Jun. 10, 2022

Accepted: --

Published Online: Sep. 29, 2022

The Author Email:

DOI:10.37188/CO.2022-0120

Topics

Table 1. Architecture details of proposed U-MVDNet

Table 1. Architecture details of proposed U-MVDNet

Table 2. Abbreviations of medical terms and corresponding color

Table 2. Abbreviations of medical terms and corresponding color

Table 3. Training parameters

Table 3. Training parameters

Table 4. Results of LABM encoder with different combinations of dilation rates

Table 4. Results of LABM encoder with different combinations of dilation rates

Table 5. Results of LABM encoder with different settings

Table 5. Results of LABM encoder with different settings

Table 6. Results of encoder with different depths when the input size is 512 × 512

Table 6. Results of encoder with different depths when the input size is 512 × 512

Table 7. Results of FFM decoder with different components

Table 7. Results of FFM decoder with different components

Table 8. Effect of dilation of U-MVDNet on mIoU

Table 8. Effect of dilation of U-MVDNet on mIoU

Table 9. Experimental results on MVD test set

Table 9. Experimental results on MVD test set

Table 10. Experimental results on ISIC 2016 + PH2 test set

Table 10. Experimental results on ISIC 2016 + PH2 test set

Table 11. Two different hardware environments

Table 11. Two different hardware environments

Table 12. Test results by different methods with different resolutions

Table 12. Test results by different methods with different resolutions