PainSeeker: a head pose-invariant deep learning method for assessing rat's pain by facial expressions.

Abstract

INTRODUCTION: Automated assessment of pain in laboratory rats is important for both animal welfare and biomedical research. Facial expression analysis has emerged as a promising non-invasive approach for this purpose. METHODS: An openly available dataset, RatsPain, was constructed, comprising images of facial expressions taken from six rats undergoing orthodontic treatment. Each image was carefully selected from pre- and post-treatment videos and annotated by eight expert pain raters using the Rat Grimace Scale (RGS). To achieve automated pain recognition, a head-pose-invariant deep learning model, PainSeeker, was developed. This model was designed to identify local facial regions strongly related to pain and to effectively learn consistently discriminative features across varying head poses. RESULTS: Extensive experiments were conducted to evaluate PainSeeker using the RatsPain dataset. After assessing the pain conditions of each rat through facial expression analysis, all tested methods achieved good performance in terms of F-score and accuracy, significantly outperforming random guessing and providing empirical evidence for the use of facial expressions in rat pain assessment. Moreover, PainSeeker outperformed all comparison methods, with an overall F-score of 0.7731 and an accuracy rate of 74.17%, respectively. DISCUSSION: The results demonstrate that the proposed PainSeeker model exhibits superior performance and effectiveness in automated pain assessment in rats compared with traditional machine learning and deep learning methods. This provides support for the application of facial expression analysis as a reliable tool for pain evaluation. The RatsPain dataset is freely available at https://github.com/xhzongyuan/RatsPain.