Deep Learning for Cell Detection in Neuroscience Research
MBF Bioscience is now leveraging how neurons learn in order to improve neuroscience research using microscopy. By incorporating artificial neural networks into MBF Bioscience software, we’re equipping neuroscientists with tools that characterize neuronal populations with unprecedented accuracy and anatomic specificity through entire brain volumes.
In the webinar titled, “Improved detection of c-fos labeled and pyramidal neurons using deep machine learning in NeuroInfo,” Dr. Gerfen, joined by Dr. Brian Eastwood and Dr. Nathan O’Connor, demonstrates how NeuroInfo uses deep learning neuronal networks to successfully detect pyramidal neurons in multiple brain regions.
Dr. Gerfen nicely summarizes the starting point of the cell detection algorithm: “The way to think about it, is that it’s a topographic map of a mountain range. Things that are very bright have very high intensity levels (e.g., peaks) relative to their surrounding areas. That’s what this software picks out.”
With that criterion as a starting point, all detected objects–which can include somas, axons, dendrites, and fiber bundles–are passed into the neural network which has been trained to identify pyramidal cell bodies. According to Dr. O’Connor, over 90 thousand labeled images were used to teach NeuroInfo’s neural network how to recognize pyramidal neurons. The end result is a neural network that contains millions of parameters able to distinguish pyramidal neurons from all the other features in an image.
“We automatically adjust millions of parameters during network training,” says computational neuroscientist Dr. Eastwood. “The training process updates these parameters to improve how the neural network identifies pyramidal cells in images. During training, we know the answers. So, we use that knowledge to inform the network about true and false outcomes to update network parameters.”. Neural networks become “smarter” and more refined as they encounter more training data. The MBF Bioscience development team leverages this to create increasingly robust neural networks for detecting a variety of cell and tissue types as we work with data from laboratories.
“The networks are getting better and smarter as we provide more types of neurons from different images of brain sections into the training sets, so that the networks are able to identify not only different types of neurons but distinguish labeled neurons from histologic processing artifacts.” says Dr. Gerfen.
You can try this advanced AI method for cell detection in our NeuroInfo software. If you have any questions about deep learning in neuroscience, email us at email@example.com, or join the discussion at forums.mbfbioscience.com.