Customizable Artificial Intelligence for the Biomedical Masses: Development of a User-Friendly Automated Machine Learning Platform for Biology Image Analysis.

摘要 研究人员和病理学家对生物医学图像的手动分析是耗时的，需要密集的 培训，并容易引入偏见和错误。组织样本、培养物或组织中靶点的光学分析 标本是检测生物特性的基础。无意偏见和注意力限制 在生物标志物的分析过程中，可能导致生物医学研究中发现的重现性差， 潜在地将错误引入临床诊断。这些问题是提供最大的 有益的循证医学，开发有效的医疗方法，提高公众信心 在科学探究中。 将计算机视觉应用于细胞目标检测是减少人类偏见的一种有前途的方法， 主观性和错误，限制了研究的可重复性，减缓了有效医疗技术的发展。 治疗。我们的图像分析软件Pipsqueak ProTM和我们的底层人工智能（AI） 技术，显著提高了组织样本分析的评价者间和评价者内的可靠性， 减少多重生物标志物的分析时间。我们的预训练细胞检测模型识别多个 细胞形态和目标类型，并实现快速，准确的图像分析，大大超过人类 分析.虽然使用预先训练的深度学习模型减少了计算和专业知识， 当分析偏离的图像时， 训练参数。 在这里，我们建议开发方法，这将大大提高机器学习的可访问性 用于跨不同领域和应用的生物医学图像分析。我们的计算机视觉服务将 通过我们的Pipsqueak Pro软件和第三方提供给研究和临床最终用户 产品集成。为了实现这些目标，我们将在先前SBIR第一和第二阶段进展的基础上， 开发了用于生物标志物检测的预训练ML模型。我们建议开发一个机器学习平台 能够减少生物医学研究和医疗保健中的人为偏见、主观性和错误， 高度创新、适应性强的“AutoML”系统。这个专利系统将允许用户轻松生成 在“无代码”环境中自定义计算机视觉功能。这里提出的具体创新将 提高强大的计算机视觉技术用于生物医学图像分析的可访问性， 让缺乏生物信息学专业知识的用户民主化地访问机器学习，深度机器学习 学习或计算机科学。从这里提出的工作产生的软件和工具将有利于 开发新的循证药物和开发有效的医疗方法。

ABSTRACT Manual analysis of biomedical images by researchers and pathologists is time consuming, requires intensive training, and is prone to introduce bias and error. Optical analysis of targets within tissue samples, cultures, or specimens is fundamental to detecting biological properties. Unintentional bias and attentional limitations during analysis of biomarkers can underlie poor reproducibility of findings in biomedical research and potentially introduce errors to clinical diagnostics. These problems are significant barriers to delivering the most beneficial evidence-based medicine, developing effective medical treatments, and promoting public confidence in scientific inquiry. Application of computer vision for cellular target detection is a promising approach to reducing human bias, subjectivity, and errors that limit the reproducibility of research and slow the development of effective medical treatments. Our image analysis software, called Pipsqueak ProTM, and our underlying artificial intelligence (AI) technology, have significantly increased inter- and intra-rater reliability of tissue sample analysis and decreased analysis time for multiplexed biomarkers. Our pre-trained cell detection models identify multiple cellular morphologies and target types and enable fast, accurate image analysis that greatly exceed human analysis. While the use of pre-trained deep learning models reduces computational and expertise requirements, detection accuracy and precision are significantly reduced when analyzing images that deviate from training parameters. Here, we propose to develop methods that will dramatically increase accessibility of machine learning for biomedical image analysis across diverse fields and applications. Our computer vision service will be made available to research and clinical end-users through our Pipsqueak Pro software and through 3rd party product integrations. To achieve these goals, we will build on our previous SBIR Phase I & II progress that developed pre-trained ML models for biomarker detection. We propose to develop a machine learning platform that is capable of reducing human bias, subjectivity, and errors in biomedical research and healthcare through a highly-innovative, adaptable “AutoML” system. This patented system will allow users to easily generate custom computer vision capabilities in a “no-code” environment. The specific innovations proposed here will improve the accessibility of powerful computer vision techniques for biomedical image analysis, by democratizing access to machine learning by users who lack expertise in bioinformatics, deep machine learning, or computer science. The software and tools resulting from the work proposed here will benefit the development of novel evidence-based medicines and development of effective medical treatments.