Smartphone-based wound infection screener and care recommender by combining thermal images and photographs using deep learning methods

使用深度学习方法结合热图像和照片，基于智能手机的伤口感染筛查和护理推荐

• 批准号: 10689270
• 负责人: Emmanuel Agu
• 金额: $ 62.85万
• 依托单位: WORCESTER POLYTECHNIC INSTITUTE
• 依托单位国家: 美国
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-09-01 至 2026-05-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10689270
• 关键词: Accident and Emergency department; Affect; Agreement; Algorithms; Amputation; Appointment; Bacteria; Blood Tests; Caregivers; Caring; Cellular Phone; Classification; Clinic; Data Set; Debridement; Detection; Diabetic Foot Ulcer; Diagnosis; Emergency Care; Emergency medical service; Excision; Home; Home visitation; Hospitals; Image; Image Analysis; Infection; Lead; Lighting; Machine Learning; Manuals; Measures; Medicare; Methods; Modeling; Operative Surgical Procedures; Patients; Pattern; Performance; Persons; Positioning Attribute; Process; Quality of Care; Quality of life; Recommendation; Recurrence; Redness; Reporting; Research; Resolution; Running; Sensitivity and Specificity; Sepsis; Services; Site; Specificity; Standardization; System; Temperature; Testing; Thermometers; Time; Tissues; Transportation; Trauma; Ulcer; Venous; Visit; Visiting Nurse; Visual; Work; Wound Infection; accurate diagnosis; beneficiary; chronic wound; clinical application; clinical decision support; cost; deep learning; detector; diabetic; digital health; evidence base; evidence based guidelines; foot; healing; improved; infection risk; learning strategy; limb amputation; machine learning algorithm; novel; point of care; pressure; response; smartphone application; standardized care; success; validation studies; wasting; wound; wound care; wound healing; wound treatment

I. PROJECT SUMMARY: Smartphone-based wound infection risk screener and care recommender by combining thermal images and photographs using deep learning methods Chronic wounds, which affect 6.5 million patients in the US12 severely affect their quality of life, can take up to a year to heal and re-occur in 60-70% of patients. Wounds often get infected (bacteria in wound), resulting in limb amputations if not treated properly and on time1. In current practice, at the Point of Care (POC) (e.g., nurses visiting patients’ homes and trauma sites), caregivers who are not wound experts have no way to diagnose infections. Thus, they cautiously refer wounds suspected to be infected to clinics for debridement of dead tissues, blood tests and infection diagnoses by experts57-60. However, referrals increase time before infected wounds are treated, and the chances of limb amputation. Moreover, some referred wounds end up not being infected, wasting patient and expert time and expenses (e.g., transportation)15-16. What is needed is a digital health solution that enables non-expert wound caregivers to accurately detect infected wounds at the POC even without debridement and provide standardized recommendations on evidence-based care and when to refer. Smartphones equipped with high resolution cameras and the processing power to run machine/deep learning methods are owned by most wound caregivers in the US56. Prior work by Goyal et al1 reported preliminary results that show that infection can be detected from visual attributes such as increased redness in/around the wound in a photograph using deep learning (accuracy 0.727± 0.025, sensitivity 0.709 ± 0.044, specificity 0.744 ± 0.05). While promising, their results need to be improved and validated before clinical applications. Moreover, their dataset included already debrided wounds with easily discernable infection cases, and they did not recommend evidence based best care and decide when referrals to wound clinics were the best course of action. Certain thermal image patterns are reliable indicators of wound infection20, and some models of smartphones are now equipped with thermal cameras55. Our hypotheses are that 1) the accuracy of smartphone wound infection detection can be improved by combining thermal images with photographs jointly analyzed using a deep learning method 2) recommendations for actionable, evidence-based wound care and when to refer can be generated using machine learning to standardize care provided by non-experts. In response to NOT-EB-19-018, we propose research to investigate the capability and accuracy of detecting infected wounds before debridement using deep learning methods applied to combinations of wound photographs and thermal images and generating care and referral recommendations. We also propose integration of the smartphone-based infection screener into our group’s existing wound assessment system7-9, 21-29 and validating it on new patients (N=100). Success on our proposed aims will increase the number and objectivity of wound infections detected outside the wound clinic and fast-tracked to the clinic for treatment, reducing the number of patients who require amputations. Our findings will apply to diverse wound types including diabetic, pressure, arterial, venous, surgical61 and trauma wounds62, which all get infected.

I.项目摘要：基于智能手机的风感染风险筛选器和护理 通过使用深度学习方法组合热图像和照片来推荐 慢性伤口影响12 us2的650万患者严重影响其生活质量，可能需要 在60-70％的患者中治愈和重新发生的一年。伤口经常被感染（细菌中的细菌），导致肢体 截肢（如果无法正确处理，则在时间1上进行处理。在目前的实践中，在护理点（POC）（例如，护士 探访患者的房屋和创伤场所），没有受伤的专家的护理人员无法诊断 感染。那就是他们灾难性地提到被感染的诊所，以调试死组织， 专家57-60的血液检查和感染诊断。但是，转介在感染伤口之前增加了时间 经过处理，以及肢体截肢的机会。而且，一些转介的伤口最终没有被感染，浪费 患者和专家的时间和费用（例如，运输）15-16。需要的是数字健康解决方案 这使得非专业伤口护理人员即使没有 清理术并提供有关循证护理以及何时参考的标准化建议。 配备高分辨率相机的智能手机和运行机器/深度学习的处理能力 方法是56 US的大多数伤口护理人员所有。 Goyal等人的先前工作报告了初步结果 这表明可以从视觉属性中检测到感染，例如伤口周围/周围的发红 在使用深度学习的照片中（准确性0.727±0.025，灵敏度0.709±0.044，特异性0.744±0.05）。 在承诺的同时，在临床应用之前，需要改进和验证其结果。而且，他们 数据集包含已调试的伤口，并带有易于辨别的感染病例，他们不建议 基于证据的最佳护理，并决定何时转诊到伤口诊所是最佳行动方案。 某些热图像模式是伤口感染的可靠指标，以及一些智能手机模型 现在配备了热摄像头55。我们的假设是1）智能手机伤口的准确性 通过将热图像与共同分析的照片结合在一起，可以改善感染检测 使用深度学习方法2）针对可行的，基于证据的逻辑护理的建议以及何时 可以使用机器学习来生成参考，以标准化非专家提供的护理。 为了响应非EB-19-018，我们提出了研究以研究检测能力和准确性 使用深度学习方法在清创术前受感染的伤口，应用于伤口的组合 照片和热图像以及产生护理和推荐建议。我们也建议 将基于智能手机的感染筛选器集成到我们小组现有的伤口评估系统7-9， 21-29并对新患者进行验证（n = 100）。在我们建议的目标上取得成功将增加数量和 伤口感染在伤口诊所外检测到的伤口感染的客观性，并快速跟踪到诊所进行治疗， 减少需要截肢的患者数量。我们的发现将适用于潜水员伤口类型 包括糖尿病，压力，动脉，静脉，手术61和创伤伤口62，它们都被感染。