Skip to main content
QUICK REVIEW

[论文解读] MAIScope: A low-cost portable microscope with built-in vision AI to automate microscopic diagnosis of diseases in remote rural settings

Rohan Sangameswaran|arXiv (Cornell University)|Aug 12, 2022
Digital Imaging for Blood Diseases被引用 6
一句话总结

MAIScope 是一种低成本、便携式显微镜,集成嵌入式视觉人工智能,利用树莓派和定制的微珠显微镜,在偏远农村地区实现疟疾自动诊断。其基于设备端的 TensorFlow Lite 模型对疟原虫感染红细胞的分类准确率达到 89.9%,可在无专业人员或稳定网络连接的环境下实现可靠、离线诊断。

ABSTRACT

According to the World Health Organization(WHO), malaria is estimated to have killed 627,000 people and infected over 241 million people in 2020 alone, a 12% increase from 2019. Microscopic diagnosis of blood cells is the standard testing procedure to diagnose malaria. However, this style of diagnosis is expensive, time-consuming, and greatly subjective to human error, especially in developing nations that lack well-trained personnel to perform high-quality microscopy examinations. This paper proposes Mass-AI-Scope (MAIScope): a novel, low-cost, portable device that can take microscopic images and automatically detect malaria parasites with embedded AI. The device has two subsystems. The first subsystem is an on-device multi-layered deep learning network, that detects red blood cells (RBCs) from microscopic images, followed by a malaria parasite classifier that recognizes malaria parasites in the individual RBCs. The testing and validation demonstrated a high average accuracy of 89.9% for classification and average precision of 61.5% for detection models using TensorFlow Lite while addressing limited storage and computational capacity. This system also has cloud synchronization, which sends images to the cloud when connected to the Internet for analysis and model improvement purposes. The second subsystem is the hardware which consists of components like Raspberry Pi, a camera, a touch screen display, and an innovative low-cost bead microscope. Evaluation of the bead microscope demonstrated similar image quality with that of expensive light microscopes. The device is designed to be portable and work in remote environments without the Internet or power. The solution is extensible to other diseases requiring microscopy and can help standardize automation of disease diagnosis in rural parts of developing nations.

研究动机与目标

  • 解决非洲和东南亚等地区农村及资源匮乏地区训练有素的疟疾显微镜技师严重短缺的问题。
  • 通过开发一种便携式、低功耗的诊断设备,克服传统显微镜存在的高成本、需电力支持和主观性强等局限。
  • 通过设备端人工智能推理实现实时、准确的疟疾点诊检测,无需互联网连接,实现在偏远地区的可靠应用。
  • 构建一个可扩展、可扩展的平台,用于诊断疟疾以外的其他显微疾病,使用相同的硬件和人工智能处理流程。

提出的方法

  • 将树莓派 4 与 10 倍变焦摄像头及触摸屏显示器集成,构成紧凑、电池供电的诊断设备。
  • 设计并实现一种低成本玻璃微珠显微镜,其成像质量与传统光学显微镜相当。
  • 部署两阶段深度学习流程:首先使用 BCCD 数据集训练血细胞检测模型,识别红细胞、白细胞和血小板;其次使用疟疾数据集训练疟原虫分类器。
  • 利用 TensorFlow Lite 对模型进行优化,实现在低端硬件上的高效推理,减小模型体积的同时保持高准确率。
  • 在有网络时实现图像上传和模型再训练的云同步功能,提升长期诊断准确率。
  • 通过简洁的用户界面实现图像采集和分析的自动化,最大限度减少用户操作,确保非专业医护人员也能轻松使用。

实验结果

研究问题

  • RQ1低成本、便携式显微镜结合设备端人工智能,能否在资源有限的农村环境中实现高诊断准确率?
  • RQ2定制微珠显微镜在疟疾诊断中,其成像质量在多大程度上可与传统光学显微镜媲美?
  • RQ3轻量级、设备端的深度学习模型在树莓派 4 等低功耗硬件上运行时,性能表现如何?
  • RQ4该系统能否在无互联网或稳定电源的离线环境中可靠运行,确保偏远诊所的可及性?
  • RQ5该系统在多大程度上可扩展至疟疾以外的其他显微疾病诊断?

主要发现

  • MAIScope 设备使用设备端 TensorFlow Lite 模型对疟原虫感染红细胞的分类平均准确率达到 89.9%。
  • 血细胞检测模型在模型尺寸显著更小的情况下,性能优于三种在高端硬件上训练的其他模型,证明其在低资源设备上的高效性。
  • 微珠显微镜产生的成像质量与昂贵的传统光学显微镜相当,验证了其在低成本诊断系统中的适用性。
  • 系统仅依靠 5V 电池组即可在离线状态下稳定运行,适用于偏远、无电网的农村诊所部署。
  • 该设备在扩展至其他需要显微分析的疾病诊断方面展现出潜力,例如组织样本中的癌症检测。
  • 未来关键步骤是开展实物血涂片的实地测试,因当前验证基于图像数据集,受限于真实血涂片的获取。

更好的研究,从现在开始

从阅读论文到最终审阅,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。