【医学+深度论文：F34】2018 CVPR Retinal Optic Disc Segmentation using Conditional Generative Adversarial

34

2018 CVPR

Retinal Optic Disc Segmentation using Conditional Generative Adversarial Network

dataset	-	introduce	github
Grishti-GS1	2017 keras	F21 复现	https://github.com/abhinav-iiit/fundus-image-segmentation
Grishti-GS1	2017 CVPR keras	F21	https://github.com/seva100/optic-nerve-cnn
-	pytorch	retinal-cGAN	https://github.com/shuangyueliao/retinal-cGAN

Method ： 分割 OD
Dataset： DRISHTI GS1
     RIM-ONE
Architecture： cGAN
Results:

Dataset	Accuracy	Dice	JACC	Senstivity	Specificity
DRISHTI	GS1	0.98	0.97	0.96	0.98
RIM-ONE	0.98	0.98	0.93	0.98	0.99

Method

The Proposed cGAN network

• cGANs is a deep learning network that can learn the statistical invariant features (texture, color etc.) of input image and segment the optic disc region.

• generator
  learns the mapping from the input, a fundus image,to the output, a segmented image.

• discriminator
  learns a loss function to train this mapping by comparing the ground-truth and the predicted output

• the whole cGAN network
  optimizes a loss function that combines a conventional binary cross-entropy loss with an adversarial term.
  The adversarial term encourages the generator to produce output that cannot be distinguished from ground-truth ones.

generator

• based on encoding and decoding layers
  LeakyRelu（slope 0.2）
  image 256×256
  

discriminator

• 5conv （3 × 3 stride 2）
  output 30×30 512
• the concatenation of the retinal image and the segmentation mask as an input to be evaluated as real or fake
  在生成器的损失计算中包含对抗性 score 有助于增强网络分割的能力

train

epoch 200 ， batch size 4 ， Adam 0.0002

Result

Experience

• 和 FCN、SegNet、U-Net 比较

• 和 论文 F14 等 共三篇文章做比较

• cGAN 效果最好,OD分割更接近地面真实情况，边界更精确。U-Net也提供了可接受的分割。在五种测试方法中，SegNet的分割效果最差。

Discussion

• cGAN网络不需要大量的图像来训练
• 由于最后的分割只在生成器网络中实现，因此在不增加任何复杂度的情况下，它可以获得很高的分割性能。
• cGAN算法优于目前最先进的 OD 分割方法