1
1. Detailed network structure for 3D Faster
R-CNN with Deep 3D Dual Path Net in
Nodule Detection
The encoder network is adapted from DPN92 directly by
changing 7×7 filters into 3×3 [1]. The numbers of blocks
are changed from 3, 4, 20, 3 to 2, 2, 2, 2. The decoder
network is to make the network symmetric. The stride 2 of
3D convolution is added in the first 3×3×3 convolution in
each block.
Stage
Output
Weights
Pre-dual
path
96×96×96, 24
3×3×3, 24
Dual path
block 1
48×48×48, 48
1×1×1, 24
3×3×3, 24, (stride 2) ×2
1×1×1, 32
Dual path
block 2
24×24×24, 72
1×1×1, 48
3×3×3, 48, (stride 2) ×2
1×1×1, 56
Dual path
block 3
12×12×12, 96
1×1×1, 72
3×3×3, 72, (stride 2) ×2
1×1×1, 80
Dual path
block 4
6×6×6, 120
1×1×1, 96
3×3×3, 96, (stride 2) ×2
1×1×1, 104
Deconv. 1
12×12×12, 216
2×2×2, 216
Dual path
block 5
12×12×12, 152
1×1×1, 128
3×3×3, 128 ×2
1×1×1, 136
Deconv. 2
24×24×24, 224
2×2×2, 152
Dual path
block 6
24×24×24, 248
1×1×1, 224
3×3×3, 224 ×2
1×1×1, 232
Output
24×24×24, 3×5
Dropout, p=0.5
1×1×1, 64
1×1×1, 15
2. Detailed network structure for 3D Faster
R-CNN with Deep 3D Residual Network
in Nodule Detection
The encoder network is adapted from Res18 directly by
changing 7×7 filters into 3×3 [2]. We find the latest
reference for 3D Res18 network in [3], and will add it into
the reference.
Stage
Output
Weights
Pre-
Residual
96×96×96, 24
3×3×3, 24
3×3×3, 24
Residual
block 1
48×48×48, 32
3×3×3, 32
3×3×3, 32, (stride 2) ×2
Residual
block 2
24×24×24, 64
3×3×3, 64
3×3×3, 64, (stride 2) ×2
Residual
block 3
12×12×12, 64
3×3×3, 64
3×3×3, 64, (stride 2) ×3
Residual
block 4
6×6×6, 64
3×3×3, 64
3×3×3, 64, (stride 2) ×3
Deconv. 1
12×12×12, 128
2×2×2, 64
Residual
block 5
12×12×12, 64
3×3×3, 64
3×3×3, 64 ×3
Deconv. 2
24×24×24, 128
2×2×2, 64
Residual
block 6
24×24×24, 64
3×3×3, 64
3×3×3, 64 ×3
Output
24×24×24, 3×5
Dropout, p=0.5
1×1×1, 64
1×1×1, 15
3. Comparison with different methods for
each fold and average false positives on
LUNA16 dataset
Supplementary Material for Deep 3D Dual Path Nets for Automated Pulmonary
Nodule Detection and Classification
Wentao Zhu
University of California, Irvine
wentaoz1@ics.uci.edu
Chaochun Liu
Baidu Research
liuchaochun@baidu.com
Wei Fan
Tencent Medical AI Lab
davidwfan@tencent.com
Xiaohui Xie
University of California, Irvine
xhx@ics.uci.edu