Kim Sueun, Yamagishi Norio, Ishikawa Shingo, Tsuchiaka Shinobu
Laboratory of Large Animal Clinical Medicine, Graduate School of Veterinary Sciences, Osaka Metropolitan University, Osaka, Japan.
BMC Vet Res. 2025 Jul 15;21(1):468. doi: 10.1186/s12917-025-04919-1.
Non-invasive temperature measurement using infrared cameras has become increasingly important for monitoring physiological changes and stress responses in animals, offering advantages over traditional rectal thermometry. However, previous methods often suffered from limitations such as environmental interference, instantaneous measurement, and inaccurate region of interest (ROI) selection due to manual settings. To overcome these limitations, studies have combined infrared cameras with AI-based segmentation to enable accurate ROI detection and to capture temporal temperature change patterns in cattle. Furthermore, the interpretability of eye and muzzle temperature measurements can vary depending on which subregions are analyzed, as areas with richer vascularization tend to display more representative temperature characteristics. To address these issues, the present study applied AI-based segmentation to infrared thermography and focused on the analysis of high-temperature, vascularized subregions within the eyes and muzzles of calves. By doing so, we aimed to enhance the clarity and reliability of temperature change pattern analysis for non-invasive monitoring of physiological status in cattle.
Thermal images were captured using a mobile infrared camera, and video recordings were obtained simultaneously from 11 calves. AI-based segmentation, utilizing previously trained weights, was used to automatically extract eye and muzzle ROIs from video images. 33 imaging sessions where the majority of frames exhibited reliable segmentation were selected for analysis. In Experiment 1, temperature data corresponding to the mean, top 10%, and top 30% values within each ROI underwent preprocessing steps (outlier rejection, standardization, and low-pass filtering) to derive temperature change patterns. This process generated six patterns per session (three for eyes and three muzzle regions), yielding a total of 198 patterns across all 33 image sessions. Cosine similarity analysis was then applied to quantify similarity within the same session. In Experiment 2, each ROI was divided into a 3 × 3 grid to map the distribution of high temperature values for spatial analysis. Statistical analyses included Kruskal-Wallis tests with Bonferroni corrections to assess regional differences.
In Experiment 1, for the eyes, the patterns derived from the top 10% and 30% of temperatures had high cosine similarity (0.94). In contrast, the patterns based on the mean values had relatively lower similarities with the top 10% and 30% patterns (0.81 and 0.86, respectively). A similar trend was observed for the muzzle: the top 10% and 30% patterns had a high cosine similarity (0.93), while the patterns based on the mean values showed lower similarities (0.80, and 0.86). In Experiment 2, for the eyes, the top 10% of temperature values were mainly in the bottom region. In comparison, the top 30% of values were more evenly distributed in the mid and bottom regions. For the muzzles, the top 10% of temperature values were mainly distributed in both the top and bottom regions, and the top 30% of values were concentrated in the mid region.
This study demonstrates that integrating AI-based segmentation with infrared thermography enables precise identification of thermally reliable subregions within the eyes and muzzles of calves, leading to the extraction of temperature change patterns with high temporal consistency. The top 10% and 30% temperature values within these regions show higher pattern similarity than mean values, with distinct spatial distributions reflecting underlying vascular anatomy. Focusing on these high-temperature, vascularized subregions enhances the interpretability and reliability of temperature change pattern analysis for non-invasive monitoring of stress and physiological status in cattle, contributing to enhanced animal welfare.
使用红外热像仪进行非侵入式体温测量对于监测动物的生理变化和应激反应变得越来越重要，相较于传统的直肠测温法具有诸多优势。然而，以往的方法常常存在局限性，如环境干扰、即时测量，以及由于手动设置导致的感兴趣区域（ROI）选择不准确。为克服这些局限性，一些研究将红外热像仪与基于人工智能的分割技术相结合，以实现准确的ROI检测，并捕捉牛的体温随时间的变化模式。此外，眼睛和口鼻部温度测量的可解释性可能因所分析的子区域不同而有所差异，因为血管分布更丰富的区域往往呈现出更具代表性的温度特征。为解决这些问题，本研究将基于人工智能的分割技术应用于红外热成像，并着重分析犊牛眼睛和口鼻部内高温、血管化的子区域。通过这样做，我们旨在提高用于牛生理状态非侵入式监测的温度变化模式分析的清晰度和可靠性。
使用移动红外热像仪采集热图像，并同时从11头犊牛获取视频记录。利用先前训练的权重，基于人工智能的分割技术用于从视频图像中自动提取眼睛和口鼻部的ROI。选择了33个成像时段进行分析，其中大多数帧呈现出可靠的分割效果。在实验1中，对每个ROI内对应于平均值、前10%和前30%值的温度数据进行预处理步骤（异常值剔除、标准化和低通滤波），以得出温度变化模式。此过程每个时段生成六种模式（眼睛三种，口鼻部三种），在所有33个图像时段中总共产生198种模式。然后应用余弦相似度分析来量化同一段内的相似度。在实验2中，将每个ROI划分为3×3网格，以绘制高温值的分布进行空间分析。统计分析包括带有Bonferroni校正的Kruskal-Wallis检验，以评估区域差异。
在实验1中，对于眼睛，由前10%和30%温度得出的模式具有较高的余弦相似度（0.94）。相比之下，基于平均值的模式与前10%和30%模式的相似度相对较低（分别为0.81和0.86）。口鼻部也观察到类似趋势：前10%和30%模式具有较高的余弦相似度（0.93），而基于平均值的模式相似度较低（0.80和0.86）。在实验2中，对于眼睛，前10%的温度值主要位于底部区域。相比之下，前30%的值在中部和底部区域分布更为均匀。对于口鼻部，前10%的温度值主要分布在顶部和底部区域，前30%的值集中在中部区域。
本研究表明，将基于人工智能的分割技术与红外热成像相结合，能够精确识别犊牛眼睛和口鼻部内热可靠的子区域，从而提取出具有高时间一致性的温度变化模式。这些区域内前10%和30%的温度值显示出比平均值更高的模式相似度，其独特的空间分布反映了潜在的血管解剖结构。关注这些高温、血管化的子区域可提高用于牛应激和生理状态非侵入式监测的温度变化模式分析的可解释性和可靠性，有助于提升动物福利。
