DeepMapi: A Fully Automatic Registration Method for Mesoscopic Optical Brain Images Using Convolutional Neural Networks
Authors: Hong Ni, Zhao Feng, Yue Guan, Xueyan Jia, Wu Chen, Tao Jiang, Qiuyuan Zhong, Jing Yuan, Miao Ren, Xiangning Li, Hui Gong, Qingming Luo, Anan Li*
Correspondence: aali@mail.hust.edu.cn
DeepMapi is a supervised deep learning based on convolutional neural network to predict the deformation field corresponding to each pair of images for registering mesoscopic micro-optical imaging datasets to the reference atlas automatically.

Environment
DeepMapi is established by using the pytorch framework and it can be used in Linux and Windows system. You need to install the following tools to start your work.

• Anaconda (Python 3.6), add to your path
• Pytorch (conda install pytorch)
• Advanced Normalization Tools (ANTs)
• conda install numpy mkl
  pip install scipy
  pip install numpy
  pip install SimpleTIK

Codes:
Download
Unzip

Data:
Download
Unzip and overwrite the 'data' folder under Codes folder

Instructions
Group you datasets

• Enter the 'data' folder, which stores the datasets of large and small deformation training set respectively.
• Taking the 'data_large_deformation' folder as an example, you need to store the datasets by number, such as 10001, 10002 and etc
• Continue to 10001_1, which is used to save the moving image and the corresponding displacement field.
• 10001_2, 10001_3, ..., are used to save the augmented dataset, you need to name you files according to the file naming method in 10001_1 folder.
• The 'data_small_deformation' folder is organized in the same format as the 'data_large_deformation' folder. You need to use the prediction results of first level to prepare you small deformation dataset.

Training and prediction

• Enter the 'code' folder, which stores the codes for the first and second level.
• Taking the 'first level' folder as an example, run according to its naming rules.
  python step1_stride_training.py
  python step2_stride_predicion.py
  python step3_feedback_network.py
  python step4_feedback_predict.py
• You can use the model of feedback network to predict the small deformation datasets, and then perform a manual registration. After that, organize your dataset by following the 'Group you dataset' section.
• Run the codes in the 'second level' folder.

Using tools
We provide the codes to evaluate the accuracy and performance of DeepMapi in 'tools' folder.