Search
Presentation Details
| Poster #230 Accuracy and Temporal Precision of Open-Source Machine Learning Models for Lick Detection |
| Georgia R. Davis1,2, Mia B. Fox1, Max L . Fletcher1, John D. Boughter Jr.1 1Department of Anatomy and Neurobiology, University of Tennessee Health Science Center , Memphis, TN, United States 2Undergraduate, Rhodes College, Memphis, TN, United States |
Lickometers using photobeams or capacitance circuits have been developed for the analysis of consummatory behaviors. While these techniques can detect licks with high accuracy and temporal precision, they lack the ability to track other orofacial behaviors which provide valuable insight into taste responses. We focused on the precision and accuracy of lick detection through video analysis from supervised machine learning models trained via open-source software. We used DeepLabCut (DLC), a pose-estimation software that is supported by Simple Behavioral Analysis (SimBA), to classify orofacial behaviors using metrics derived from pose information. Head-fixed, water restricted C57Bl/6 mice were placed on a treadmill and trained to lick a spout after a brief tone. Videos of the ventral aspect of the face were acquired either at 30, 60, or 160 fps; respiration was detected with an external thermistor wire placed in front of one nostril. Lick detection accuracy was determined by comparing the output of a capacitance circuit lickometer to DLC pose estimation of tongue location and a region of interest (spout), and a behavior classification model trained in SimBA on spout licks. Temporal precision was determined through the lens of respiratory phase preference of licks using thermistor data collected at 25k Hz and down-sampled to 2000 Hz. Analyses were performed on videos at each frame rate. While the accuracy of lick counts appears unaffected by frame rate, temporal precision of lick detection decreases as frame rate decreases. Behavioral analysis performed on videos acquired at 160 fps yields a similar respiratory phase preference, suggesting its temporal resolution is sufficient for detailed analysis of electrical signal acquired at 25k Hz. |