Page 207 - Computational Retinal Image Analysis

1 Introduction 203

Inverse Polar Transformation Output Up-Sample Down-Sample MaxPooling DeConv <2x2, 2> Conv <1x1, 1>, with Sigmoid Conv <3x3, 1>, with ReLU Copy and Merge Multi-label Loss

+ 400 x 400 Side-Output + Multi-label
Multi-label Map L (1) s L (2) s L (3) s L (4) s

2 2 2 2
32 400 x 400 400 x 400 400 x 400
32
64
64 400 x 400 64
Segmentation 128 128
M-Net 200 x 200 128

256 256

256 512
100 x 100
Transformation 50 x 50 512 U-Shape Convolutional Network
Polar 512

256
256 25 x 25
Optic Disc Detection 128 256 384

128

64 192
64
96
32 32 Illustration of the M-Net framework.
Fundus Image Input 3 64 3 128 3 256 Multi-scale

3
FIG. 2
50 x 50
100 x 100
400 x 400
200 x 200