Development of a deep neural network to measure spontaneous pain from mouse facial expressions

开发深度神经网络来测量小鼠面部表情的自发疼痛

• 批准号: 10579988
• 负责人: Mark J. Zylka
• 金额: $ 37.61万
• 依托单位: UNIV OF NORTH CAROLINA CHAPEL HILL
• 依托单位国家: 美国
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-02-15 至 2025-01-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10579988
• 关键词: Acetic Acids; Adopted; Adoption; Analgesics; Animal Model; Biological Assay; Body Regions; Cellular Phone; Characteristics; Classification; Color; Color blindness; Computer software; Computers; Custom; Data; Data Analytics; Database Management Systems; Dependence; Development; Face; Facial Expression; Facial Pain; Failure; Future; Genes; Goals; Human; Human Resources; Image; Laparotomy; Manuals; Maps; Measures; Medicine; Methods; Modality; Modeling; Motion; Mus; Neural Network Simulation; Non-Steroidal Anti-Inflammatory Agents; Opioid Analgesics; Output; Pain; Pain Measurement; Pain intensity; Pharmaceutical Preparations; Postoperative Pain; Pre-Clinical Model; Publishing; Reproducibility; Research Personnel; Resources; Rodent; Science; Screening procedure; Sleep; Specificity; System; Testing; Time; Training; addiction; analytical tool; associated symptom; chronic pain; cloud based; convolutional neural network; data repository; deep neural network; efficacy evaluation; handheld mobile device; improved; inflammatory pain; interest; machine learning algorithm; machine learning model; mobile application; multimodality; neural circuit; novel; open source; opioid epidemic; overdose death; pain relief; painful neuropathy; persistent symptom; pre-clinical; recurrent neural network; side effect; smartphone application; spontaneous pain; success; tool; transfer learning; user-friendly; web platform; web services

PROJECT SUMMARY Opioid analgesics are commonly used to treat pain but have serious side effects, including addiction, dependence, and death from overdose. While there is a significant need for new non-addictive analgesics, efforts to develop new pain medicines have met with limited success. In part, this failure is due to an overreliance on evoked pain measures in preclinical models. Indeed, most preclinical models do not measure spontaneous pain—the main symptom of chronic pain in humans. To increase translational relevance, the Mouse Grimace Scale (MGS) was developed to quantify characteristic facial expressions associated with spontaneous pain. The MGS is reproducible across labs and was used to evaluate the efficacy of analgesics. However, the MGS has not been widely adopted due to its high resource demands and low throughput. To overcome this limitation, we adapted a machine learning model to classify the presence or absence of pain from mouse facial expressions. We called this model the automated Mouse Grimace Scale (aMGS). After training, this model identified mice in pain with 94% accuracy, comparable to a highly-trained human. However, our original “aMGS 1.0” is limited in several respects. It is only accurate at detecting facial grimacing in white- coated mice, and produces a binary assessment (“pain” vs. “no pain”) instead of a graded score. Moreover, aMGS 1.0 cannot dynamically determine pain status from full-motion videos. Additionally, we relied on an older piece of software that does not consistently extract high-quality images of the mouse face. The aMGS 1.0 also has difficulty distinguishing between images of sleeping and grimacing mice. Finally, aMGS 1.0 suffers from a “black box” problem inherent to most machine learning algorithms, in that we do not know what facial details it uses to produce a pain assessment. Here we propose to overcome all of these limitations by developing a more sophisticated version of our automated pain classifier (aMGS 2.0). To achieve this goal we will: 1) Develop and validate a new open-source platform to classify (frame-by-frame) spontaneous pain intensity from mouse facial expressions, using albino (white) mice and motion information. 2) Enhance the generality of aMGS 2.0 for use with black mice. And, 3) Develop a user-friendly web-based platform that operates on computer-based and mobile devices. We will validate the utility of aMGS with three pain assays that produce grimaces in rodents—inflammatory pain, post-surgical (laparotomy) pain, and neuropathic pain. To increase rigor and reproducibility, two pain assays will be performed and scored with aMGS 2.0 in an independent lab. Numerous investigators in the pain field have expressed interest in using our proposed model. The platform will include a cloud-based data repository and analytic tools to facilitate curation of public data, continuous improvement of the model over time, and integration of new analytic tools. One analytic tool that we plan to develop will identify mouse features that most influence pain classification.

项目摘要 阿片类镇痛药通常用于治疗疼痛，但有严重的副作用，包括成瘾， 依赖和过量死亡。虽然对新的非成瘾性镇痛剂存在显著需求， 开发新的止痛药的努力取得了有限的成功。在某种程度上，这种失败是由于 过度依赖诱发疼痛的措施在临床前模型。事实上，大多数临床前模型并不测量 自发性疼痛-人类慢性疼痛的主要症状。为了增加翻译相关性， 小鼠鬼脸量表（MGS）被开发用于量化与以下相关的特征性面部表情： 自发性疼痛。MGS在实验室间具有重现性，用于评价镇痛剂的疗效。 然而，MGS由于其高资源需求和低吞吐量而没有被广泛采用。到 为了克服这一限制，我们采用了一种机器学习模型来对疼痛的存在或不存在进行分类， 从老鼠的面部表情。我们将此模型称为自动鼠标鬼脸量表（aMGS）。后 通过训练，该模型以94%的准确率识别出疼痛的小鼠，与训练有素的人类相当。然而，在这方面， 我们最初的“aMGS 1.0”在几个方面受到限制。它只在检测白色的面部表情时准确- 包被的小鼠，并产生二元评估（“疼痛”对“无疼痛”）而不是分级评分。此外，委员会认为， aMGS 1.0无法从全动态视频中动态确定疼痛状态。此外，我们依靠一个更老的 这款软件不能始终如一地提取高质量的鼠标面部图像。AGS 1.0还 很难区分睡觉的老鼠和做鬼脸的老鼠。最后，aMGS 1.0遭受了 大多数机器学习算法所固有的“黑箱”问题，因为我们不知道它是什么面部细节。 用于产生疼痛评估。在这里，我们建议通过开发 我们的自动疼痛分类器（aMGS 2.0）的更复杂版本。为了实现这一目标，我们将：1） 开发并验证一个新的开源平台，以分类（逐帧）自发性疼痛强度， 小鼠面部表情，使用白化病（白色）小鼠和运动信息。2)增强的通用性 aMGS 2.0用于黑小鼠。以及，3）开发一个用户友好的基于网络的平台， 基于计算机的和移动的设备。我们将用三种疼痛测定法验证aMGS的效用， 啮齿类动物的鬼脸-炎性疼痛、手术后（剖腹术）疼痛和神经性疼痛。增加 严格性和再现性，将在独立实验室中进行两次疼痛测定并使用aMGS 2.0评分。 疼痛领域的许多研究人员都表示有兴趣使用我们提出的模型。平台 将包括一个基于云的数据存储库和分析工具，以促进公共数据的管理， 随着时间的推移改进模型，并整合新的分析工具。我们计划用一种分析工具 开发将确定最影响疼痛分类的小鼠特征。