Table of Contents
Synthetic Histology: Improvement to Image Registration
We have been exploring unsupervised methods to transform between different image modalities in neuroimaging data which we call synthetic histology. One outcome of this is that we can use the results to improve image registration accuracy between different modalities. We can first translate one contrast domain into another, and then perform image registration in the same image modality space. We can also add further constraints such as cortical boundary segmentation automatically generated using the same learning framework.
We trained the unpaired image-to-image translation machine learning algorithm called CycleGAN, to translate from block face to synthetic myelin images. We then use another paired image-to-image translation algorithm (Pix2Pix) to translate from synthetic myelin images to left and right cortical segmentation images.
The above figure shows the synthetic myelin brain and its cortical segmentations after processing all slices of block face.
We then register the real myelin to the synthetic myelin with cortical segmentation as additional constraints (as shown in the above figure).
Tracer Segmentation Using A.I.
We used the U-Net deep learning architecture and trained it on manually drawn tracer masks to automatically segment anterograde tracer signals in high resolution images (up to 30,000 by 30,000 pixels). This approach was used in our NanoZoomer Connectomics Pipeline.
Video showcasing the result of our tracer segmentation using U-Net on a high resolution image of a slice of a marmoset brain injected with anterograde tracer and imaged using a NanoZoomer HT 2.0 slide scanner kindly provided by Drs Noritaka Ichinohe & Hiroshi Abe @ RIKEN CBS.