终于代码码完，基本测试做完，评估也基本弄好，开始要动笔码字了！！
但是latex模板打开，发现只有个模糊的outline，恩，先来这里整理下思绪。
master thesis 从六月初开始做，一直到11月底，当中各种墨迹，出事情，解决事情，找实习，开始实习，终于，终于在十一月底基本测试完了。结果目前来说，我还是比较满意的。
thesis的题目暂时定为：
Deep Learning for Automated Bone Age Assessment
目前完成的主要内容就是
1. 读取X-ray骨骼图片，进行医疗图片分割(unet, FCN)，去除noise部分。
2. Oversampling 一部分，使得training data 更加normalise
3. 因为是灰度图片，但是又要使用pretrained weights，所以换成3 channels, 进行CLAHE 变换，再进行adaptive equalisation
4. 数据集10000左右，进行data augmentation（这块比较重要）
5. transfer learning， fine tuning the network，进行regression task
6. 发展了一个新网络，spatial transformer layer + VGG，进行regression task
7. 测试，visualisation, attention map, show the network
好了，我的outline是
1. Introduction
1.1 Background
1.2 Motivation
1.3 Objectives
1.4 Related Work
1.5 Thesis organisation
2. Preliminaries on Convolutional Neural Networks
2.1 Artificial Neural Network
2.2 Convolutional neural network (convolutional layers, pooling, fully connected layer)
2.3 Supervised Training of Neural Network
2.4 Regularisation for neural network
3.5.1 Weight decay
3.5.2 Dropout
3.5.3 Batch normalisation
2.5 Transfer Learning
4. Theoritical background and Implementation
4.1 Methodology Pipeline
4.2 Image Segmentation for X-ray image
4.3.1 Preprocessing
4.3.2 Data Augmentation
4.3.2 neural networks for semantic segmentation
Unet
FCN
4.3.3 post processing
4.3 Regression for bone age assessment
4.3.1 neural networks selection
4.3.2 Ensemble
4.4 Training Methods
4.4.1 Hyperparameter definition
4.4.2 K fold cross validation
5. Experimental Evaluation
5.1 Datasets (RSNA, Essen hospital)
5.2 Experimental setup (hardware?)
5.3 Image segmentation
5.3.1 Evaluation
5.3.2 Model exploration (also including noise)
5.3.3 Different Model comparison
5.4 Regression for bone age assessment
5.4.1 Evaluation
5.4.2 Different Models comparison
5.4.3 Saliency map
6. Discussion
7. Conclusion
网友评论