Analyzing 2D EM image sequences of biological tissues in a 3D context is necessary to restore the original 3D integrity destroyed by physical sectioned by aligning the serial 2D images. However, similar texture intra-section and complex variations of serial EM images intersections make it challenging to find the correct correspondences. Inspired by finding correspondences manually, which uses the topological relationship of image contents, we developed an attention-based neural network method for serial EM image registration to improve the registration accuracy. More detail can be found in the following paper:
Human Attention-inspired Volume Reconstruction Method on Serial Section Electron Microscopy Images
Our method was trained on the Pytorch deep learning framework. The required libraries are as follows.
python3.7, numpy, torch, torchvision, opencv
If you don't have some of these libraries, you can install them using pip or another package manager.
If you just want to test our method, you can use "./main.py". The detailed usage and code annotations will be added later.
Here are some examples of aligning serial EM images using different image registration algorithms. And intuitively, our method has also achieved the best results
Please refer to the paper "Human Attention-inspired Volume Reconstruction Method on Serial Section Electron Microscopy Images".