The MacqD deep-learning-based model for automatic detection of socially housed laboratory macaques

Area of Science:

• Computer Vision
• Animal Behavior Analysis
• Machine Learning

Background:

• Current video-based behavior analysis models struggle with detecting macaques in challenging laboratory environments.
• Limitations include occlusions, glass reflections, and lighting variations common in animal housing.
• There is a need for robust macaque detection systems for research applications.

Purpose of the Study:

• To develop and evaluate MacqD, a novel deep learning model for accurate macaque detection.
• To enhance attention-based feature extraction using a SWIN transformer backbone within a Mask R-CNN framework.
• To assess MacqD's performance against existing models in complex laboratory conditions.

Main Methods:

• Developed MacqD, a modified Mask R-CNN model with a SWIN transformer backbone.
• Collected and analyzed video frames from 20 caged rhesus macaques at Newcastle University.
• Compared MacqD's performance against pre-existing macaque detection models using F1-scores.

Main Results:

• MacqD achieved a median F1-score of 99% for single macaques and 90% for two macaques in focal cages.
• Generalization tests on new macaques yielded median F1-scores of 95% (single) and 81% (two).
• MacqD demonstrated strong generalization capacity with a 90% F1-score on paired macaques from another facility.

Conclusions:

• MacqD significantly outperforms existing models in detecting macaques in complex laboratory environments.
• The model exhibits robust performance under challenging conditions like occlusions and reflections.
• MacqD shows strong generalisation capabilities across different settings and macaque groups.