Three-Dimensional Parallel Convolution Neural Network Brain Tumor Segmentation Based on Dilated Convolution

Bowen Feng; Xiaoqi Lü; Yu Gu; Qing Li; Yang Liu

doi:10.3788/LOP57.141009

Laser & Optoelectronics Progress, Volume. 57, Issue 14, 141009(2020)

Three-Dimensional Parallel Convolution Neural Network Brain Tumor Segmentation Based on Dilated Convolution

Bowen Feng¹, Xiaoqi Lü^1,2,3、*, Yu Gu^1,3, Qing Li¹, and Yang Liu¹

¹Inner Mongolia Key Laboratory of Pattern Recognition and Intelligent Image Processing, School of Information Engineering, Inner Mongolia University of Science and Technology, Baotou, Inner Mongolia 0 14010, China

²School of Information Engineering, Inner Mongolia University of Technology, Hohhot, Inner Mongolia 0 10051, China

³School of Computer Engineering and Science, Shanghai University, Shanghai 200444, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(15)

Fig. 1. Image data of BraTS 2018 dataset. (a) FLAIR; (b) gold standard; (c) mask^[15-17]

Download full size

Fig. 2. Schematic diagram of dilated convolutional path

Download full size

Fig. 3. Schematic diagram of parallel CNN

Download full size

Fig. 4. Kernel of the dilated convolution. (a) Standard convolution kernel; (b) dilated convolution with filling rate of 1; (c) dilated convolution with filling rate of 3

Download full size

Fig. 5. Kernel of the jagged convolution

Download full size

Fig. 6. Structure diagram of DenseNet model

Download full size

Fig. 7. Module of dense connection

Download full size

Fig. 8. Module of transition

Download full size

Fig. 9. Average Dice coefficient of segmentation results of different deep networks

Download full size

Fig. 10. Structure of dilated convolutions and Dice coefficients of its segmentation results. (a) Schematic diagram; (b) average Dice coefficients

Download full size

Fig. 11. Evaluation index of brain tumor total segmentation results. (a)Average accuracy; (b) sensitivity index; (c) specificity index; (d) average Dice coefficient

Download full size

Fig. 12. Visual segmentation of tumor tissues by optimization model. (a) Sagittal images; (b) axial images; (c) coronal images

Download full size

Table 1. Structure of densely connected network Dense-12

View table

Table 1. Structure of densely connected network Dense-12

Layer	Path of densely connected network
Convolution	3×3×3 conv, stride 2
Pooling	3×3×3 max pooling, stride 2
Densely connected	$[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×30, $[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×30, $[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×30, $[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×30
Transition	1×1×1 conv, 2×2×2 average pooling, stride 2
Densely connected	$[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×40, $[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×40, $[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×40, $[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×40
Transition	1×1×1 conv, 2×2×2 average pooling, stride 2
Densely connected	$[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×50, $[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×50, $[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×50, $[\begin{array}{l} 1 \times 1 \times 1 conv \\ 3 \times 3 \times 3 conv \end{array}]$ ×50
Classificationlayer	3×3×3 global average pooling, fully connected, Softmax

Table 2. Comparison of CNNs with different depths
View table
Table 2. Comparison of CNNs with different depths
Layer Sensitivity Specificity Average Dice Time /h
Dense-8 0.8076 0.9603 0.6719 141.7
Dense-10 0.8315 0.9671 0.7095 189.6
Dense-12 0.8784 0.9835 0.8690 256.0
Dense-15 0.8619 0.9883 0.8836 359.5

Table 3. Comparison of segmentation results of various tumor tissues by different models
View table
Table 3. Comparison of segmentation results of various tumor tissues by different models
Model Dice coefficient
Complete Core Enhancing
Dilatedconvolution-CNN 0.90 0.73 0.71
Ref.[6] 0.88 0.79 0.73
Ref.[32] 0.88 0.87 0.81
Ref.[33] 0.87 0.81 0.78
Ref.[11] 0.90 0.76 0.73
Ref.[34] 0.88 0.83 0.77
Ref.[35] 0.90 0.85 0.81

Tools

Get Citation

Copy Citation Text

Bowen Feng, Xiaoqi Lü, Yu Gu, Qing Li, Yang Liu. Three-Dimensional Parallel Convolution Neural Network Brain Tumor Segmentation Based on Dilated Convolution[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141009

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites