As we all realize, segmentation is probably the most important step of the pipeline. Notwithstanding the quality of the registration of serial sections, the quality and reproducibility of the segmentation algorithm is the key to the success of the whole mapping process.
This page will be dedicated to comparison of some different approaches, which presently include:
- Segmentation using algorithms of Mathematical Morphology. This approach was initially developed using texture analysis system TAS (Ernst Leitz Wetzlar, GmbH) in 1980s, and recently reproduced using MAMBA Image library (CMM, France). We describe this approach not only for historical reason, but for a practical reason as well. It turned out to be very useful to start creating training and evaluation image data sets for next two approaches automatically, using the described algorithm. After visual control and manual editing of created images they can be efficiently used for training and validation of AI networks. <<link to the page>>
- Segmentation algorithm using UNET CNN in Neural Network Console (Sony), which is new and very promising application, which might become very useful, especially for beginners in the area of AI and deep learning. <<link to the page>>
- Segmentation algorithm based on Masked RCNN developed in Python using Keras and Tensor Flow libraries. Preliminary tests showed that this network delivers the most accurate results. <<link to the page will be added ASAP>>
Copyright © 2017-2018, Brain Imaging Informatics, LLC