Automated mouse behaviour recognition for neuroscience research labs

Abstract

Research studies in diverse fields, like neuroscience, usually employ experimentation on mice. Recording them is one of the most common ways of monitoring their actions and behaviours to certain stimuli. However, this videos later require professional analysis, which is a tedious and time consuming task. Current computer vision methodologies provide us with the necessary tools to build automated models that could perform this laborious and repetitive task. This project aimed to research and develop computer vision and machine learning approaches enough to perform behaviour classification at frame level on laboratory mice records. The primary objective was to develop a tool for researchers to analyze mouse behavior, alleviating the repetitive and time-consuming manual analysis. Our focus was to align our methodologies with real-world scenarios where data varies significantly. By examining state-of-the-art sequence analysis techniques, we identified key challen- ges and limitations in our data. This led to the development of tools such as frames-per-second ratio regulation and automatic mouse position detection, which improved our models’ ability to handle diverse video inputs. Through extensive experimentation and benchmarking, we designed a robust pipeline for sequence classification, achieving precision and recall rates exceeding 90% across various recording conditions. Our tool effectively adapts to different lighting, camera placements, and orientations, enhancing its applicability in real-world settings. Furthermore, we integrated these automated models with an intuitive User Interface, providing researchers with easy access to this tool.