[论文精读]U-Net: Convolutional Networks for BiomedicalImage Segmentation

论文原文：U-Net: Convolutional Networks for Biomedical Image Segmentation (arxiv.org)

英文是纯手打的！论文原文的summarizing and paraphrasing。可能会出现难以避免的拼写错误和语法错误，若有发现欢迎评论指正！文章偏向于笔记，谨慎食用！

1. 原文逐段精读

1.1. Abstract

        ①Reasonable use of annotation samples

        ②"The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization"

        ③This model is for segmenting neuronal structures in electron microscopic stacks

        ④This model peforms great in small training sample 

1.2. Introduction

        ①The expectations for machine learning and deep learning in medicine often lie not in classification accuracy, but in region segmentation and other aspects

        ②They consider the sliding-window model by Ciresan et al. as slow in training and inaccuracy brought by maxpooling

        ③⭐U-Net takes upsampling instead of pooling

        ④什么重叠贴图策略？？我没能明白，为啥这样就能预测

        ⑤They use elastic deformations to augment there data, which keeps the invariance

1.3. Network Architecture

        ①The whole framework: 

        ②3*3 convolutions include no padding

        ③Stride of maxpooling is 2

        ④Double the number of channels when downsampling

        ⑤Up-conv 2*2 halves the number of feature channels

1.4. Training

1.4.1. Data Augmentation

1.5. Experiments

1.6. Conclusion

2. 代码

3. Reference List

Ronneberger, O., Fischer, P. & Brox, T. (2015) 'U-Net: Convolutional Networks for Biomedical Image Segmentation', MICCAI 2015: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, pp 234–241. doi: U-Net: Convolutional Networks for Biomedical Image Segmentation | SpringerLink