论文阅读系列–自监督学习篇
1. DiRA: Discriminative, Restorative, and Adversarial Learning for Self-supervised Medical Image Analysis
一句话摘要: 整合了 Discriminative learning, restorative learning, and adversarial learning
论文之我问:
1.1 这个论文的动机是什么
众所周知，计算机视觉中的自监督学习有Discriminative learning, restorative learning, and adversarial learning，这三种都给自监督学习提供了很大的帮助。但是，当下研究忽视了这三者之间的协同效应(1+1+1)。
1.2 这篇论文的研究框架是什么样的
全新Loss设置: \(L = \lambda _ { d i s } * L _ { d i s } + \lambda _ { r e s } * L _ { r e s } + \lambda _ { a d v } * L _ { a d v }\)
其中:
\[L _ { \text { dis } } = \ell ( z _ { 1 } , z _ { 2 } )\\ L _ { r e s } = E _ { x } \operatorname { dist } ( x _ { 1 } , x _ { 1 } ^ { \prime } )\\ L _ { a d v } = E _ { x } [ \log D _ { \phi } ( x _ { 1 } ) ] + E _ { x } [ \log ( 1 - D _ { \phi } ( x _ { 1 } ^ { \prime } ) ) ]\]具体解释: \(\ell ( z _ { 1 } , z _ { 2 } )\) 表示距离/相似度
1.3 这篇论文的实验效果
2D对比的使用了:
- 对比模型： MoCov2, Barlow Twins, SimSiam
- 数据集: ChestXray14
- 模型架构: 2D U-net with ResNet-50
- 超参设置:
- 优化器: Adam
- $L_{res}$: MSE损失
- $D _ { \phi }$ : four convolutional layers with the kernel size of 3×3
- batch size: 256
- 设备: 4 Nvidia V100 GPUs
- $\lambda_{res},\lambda_{adv},\lambda_{dis}$ : 10, 0.001, 1
- input images: 224×224
- learning rate: 2e-4 and (β1, β2) = (0.5, 0.999)
3D对比使用了:
- 对比模型：TransVW
- 数据集: 623 chest CT scans in the LUNA
- 模型架构: 3D U-net + classification head
- 超参设置:
- 优化器: Adam
- $L_{res}$: MSE损失
- $D _ { \phi }$ : four convolutional layers with the kernel size of 3×3×3
- batch size: 256
- 设备: 4 Nvidia V100 GPUs
- $\lambda_{res},\lambda_{adv},\lambda_{dis}$ : 100, 1, 1
- epoch: 200
- learning rate: 1e-3
Tasks:
-
ChestX-ray14, CheXPert [30], SIIM-ACR [1], and NIH Montgomery [31] for 2D models,
-
LUNA, PE-CAD [41], LIDC-IDRI [2], LiTS [5], and BraTS [4] for 3D models
1.4 代码参考
https://github.com/JLiangLab/DiRA.*
