Performing a detailed behavior analysis is crucial to understanding brain behavior relationship. One of the best ways to evaluate behavior is through careful observations. However, quantifying the observed behavior is time consuming and challenging.
Classical methods of behavior analysis are not easily quantifiable and are inherently subjective. Recent developments in deep learning, a branch of machine learning and artificial intelligence fields, provide opportunities for automated and objective quantification of images and videos. Here we present our recently developed methods utilizing deep neural networks to perform detailed behavior analysis in rodents and humans.
The main advantage of this technique are its flexibility and applicability to any imaging data for behavior analysis. The DeepBehavior Toolbox supports single object identification, multi object detection, and human pose tracking. We also provide the post processing code in MATLAB for more in-depth kinematic analysis methods.
Begin by setting up Tensor Box. Activate the environment, then use GitHub to clone Tensor Box and install it on the machine and on additional dependencies. Next, launch the labeling graphical user interface and label at least 600 images from a wide distribution of behavior frames.
To label an image click the top left corner of the object of interest and then the bottom right corner. Then make sure that the bounding box captures the entire object. Click next to move to the next frame.
To link the training images to a network hyper parameters file, open overfeat_rezoom. json in a text editor and replace the file path under train_idl to labels.json. Then add the same file path under test-idl and save the changes.
Initiate the training script which will begin training for 600, 000 iterations and generate the resulting trained weights of the convolutional neural network in the output folder. Then perform prediction on new images, and view the outputs of the network as labeled images and as bounding box coordinates. Install YOLOv3.
Then label the training data with Yolo_mark by placing the images in the Yolo_mark-data-obga folder and labeling them one by one in the graphical user interface. Label approximately 200 images. Next, set up the configuration file.
To modify the configuration file open the YOLO-obj. cfg folder. Modify the batch, subdivision, and classes lines.
Then change the filter for each convolution layer before a YOLO layer. Download the network weights and place them into the darknet-build. x64 folder.
Run the training algorithm, and once it is complete view the iterations. To track multiple body parts in a human subject, install OpenPose then use it to process the desired video. The capabilities of the DeepBehavior Toolbox were demonstrated on videos of mice performing a food pellet reaching task.
Their right paws were labeled and movement was tracked with front and side view cameras. After post processing with camera calibration, 3D trajectories of the reach were obtained. The outputs of YOLOv3 are multiple bounding boxes because multiple objects can be tracked.
The bounding boxes are around the objects of interest which can be parts of the body. In OpenPose, the network detected the joint positions and after post processing with camera calibration, a 3D model of the subject was created. One critical step not covered in this protocol is ensuring that your device has the appropriate Python versions and dependencies as well as a GPU configured device before beginning.
After successfully obtaining the track behavior from the network additional post processing can be done to further analyze the kinematics and patterns of the behavior. Why the DeepBehavior Toolbox is applicable for diagnostic approaches in disease models of rodents and human subjects is not direct therapeutic benefit. Use of these techniques as a diagnostic or prognostic tool is under active research within our laboratory.
This technique is being used to investigate the neural mechanisms of skilled motor behavior in rodents as well as being used in clinical studies to evaluate motor recovery in patients with neurological diseases.
