DeepLabCut network trained to track mouse 'front' body parts during rotarod running (front-view)

DeepLabCut (https://github.com/DeepLabCut/) (Mathis et al., 2018; Nath et al., 2019) was used for tracking body parts of mice in an open field arena or in the rotarod. DeepLabCut 2.1.8.2 (local version on Windows with CPU, using the GUI) and 2.1.10.2 (google colab to train the network) were used using default parameters and the pretrained resnet50 network with imgaug augmentation. Frames were extracted with the k-means method and outlier frames with the jump method. Rotarod, front camera: 29 frames from 18 videos (10 fps) were extracted for a total of 520 labeled pictures. 12 body parts (left, right and mid snout, left/right top/bottom ears, left/right eyes, headmount, left/right foot), 4 corners of the rotarod and 4 points on the rotarod wheels were manually labeled and linked to each other using skeletons. A neural network was trained using these images for 225K iterations (train error: 1.58, test error: 1.63). 20 outlier frames were extracted from each video and relabeled. An additional 20 images from 20 new videos with different recording conditions were labeled. The network was then refined for 331K iterations (from scratch) (train error: 2.34, test error: 5.49). This process was repeated a second time when adding 20 new videos (400 frames) and the network trained to a final 402K (train error: 2.64, test error: 3.98). For this last batch, brightness/contrast were too low to detect body features; brightness/contrast were thus enhanced using custom-written Python scripts. Relevant videos were analyzed at each of the 3 steps, for a total of 152 videos.

Used to analyze videos for a publication (Labouesse et al., Nature Communications 2023).

Network not included in the final publication.