Visual Prediction of Mouse Mass using Machine Learning

Authors

Malachy Guzman

Document Type

Article

Publication Date

Summer 2022

Keywords

JMG

JAX Location

In: Student Reports, Summer 2022, The Jackson Laboratory

Abstract

Mouse model experiments often rely on body weight as a key indicator of health and as a measurable experimental effect. This makes measurement accuracy important, but the frequent handling required to weigh mice on a scale can induce physiological stress responses that potentially interfere with the intended experiment. Efforts have been made to engineer physical solutions, but these approaches are typically neither scalable nor applicable to a wide range of experimental setups. To avoid these problems, we propose a non-invasive, generalizable approach using computer vision to predict mouse mass from video data. We apply segmentation mask and ellipse-fit deep neural networks to extract information about the area and geometry of approximately 2400 mice in open field arenas across both sexes and 62 different strains. We combine this visual data with sex, strain, age, and other information to build several models with different variables for different use cases, including a visual-only model, a non-genetic model, and a full model. Over a 1-2 hours of sampling per mouse, the visual-only model achieves a mean average error of 1.81 grams. In the same conditions, our full model predicts individual mouse weight with a mean absolute error of 1.28 grams and mean relative error of 5.4% ± 5.7% between our predicted and true mass of the mouse.

Please contact the Joan Staats Library for information regarding this document.

COinS