Automatic detection and analysis of visual patterns and movement of lampreys for neurological investigation: A Machine Learning approach to calculations for a study of neurological development of visual movements

Abstract

Visual stabilization is a crucial feature that allows multiple advanced visual behaviors by clamping the visual scene, and it is performed via optokinetic and vestibular reflexes. The vestibulo-ocular reflex is well neurologically described and known to be mostly shared between vertebrates, but its evolutionary origins and how it interacts with vision are less known. The study of this integration and mechanisms is performed using lampreys, belonging to the oldest vertebrates alive, due to their more simplistic nervous system and experimental accessibility. An ongoing study shows that they exhibit robust visuo-vestibular integration. In order to obtain deeper knowledge, it is necessary to know if they exhibit saccades in the form of nystagmus, as this would show that the basic components of all eye movements were already present at the dawn of vertebrate evolution, indicating that the underlying basic neuronal pathways are conserved. For this, videos with different experimental preparations are recorded and then analyzed, not only for nystagmus, but also for every eye movement. This requires a precise analysis that, if performed manually, is very tedious and time consuming, and can lead to many errors due to the required precision of measure for each frame of the lamprey videos. In order to overcome this issue, a machine learning approach is taken. DeepLabCut is a software for deep learning using python that specializes in animal pose tracking and estimation. Trainings with different neural nets and parameters were tried to find the combination that best fitted the necessities of the group, regarding accuracy and efficiency. In addition, some python scripts were created to ease the training process for people less familiarized with the software.