Enhancing Assisted Reproductive Technologies with Deep Learning and Data Visualization

通过深度学习和数据可视化增强辅助生殖技术

• 批准号: 10376335
• 负责人: Dalit Ben Yosef
• 金额: $ 68.39万
• 依托单位: HARVARD UNIVERSITY
• 依托单位国家: 美国
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-04-01 至 2026-03-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10376335
• 关键词: Adopted; Age; Assisted Reproductive Technology; Back; Cell Lineage; Classification; Clinic; Clinical; Clinical Data; Clinical Decision Support Systems; Clinical Treatment; Cloud Service; Communities; Complex; Computer Models; Computer Vision Systems; Computers; Couples; Data; Data Analyses; Data Set; Data Storage and Retrieval; Databases; Decision Making; Detection; Development; Discipline; E-learning; Electronic Health Record; Embryo; Embryo Transfer; Embryonic Development; Fostering; Goals; Human; Human Biology; Image; Image Analysis; In Vitro; Judgment; Knowledge; Label; Lead; Machine Learning; Manuals; Medical; Methods; Modeling; Morbidity - disease rate; Morphology; Mothers; Multiple Pregnancy; Obesity; Patients; Pattern; Physiological; Pre-implantation Embryo Development; Pregnancy; Pregnancy Rate; Privacy; Probability; Research; Science; Scientist; Secure; Security; Text; Time; Trees; United States; Ursidae Family; Uterus; Visual; Visualization; Visualization software; analytical tool; base; blastocyst; clinical decision-making; clinical practice; cloud based; cohort; convolutional neural network; data cleaning; data curation; data management; data visualization; deep learning; deep learning model; embryo cell; embryo monitoring; feature extraction; human-in-the-loop; implantation; improved; infertility treatment; insight; large scale data; machine learning algorithm; machine learning model; microscopic imaging; model design; multi-task learning; multimodality; novel; operation; predictive modeling; success; supervised learning; tool; unsupervised learning; zygote

PROJECT SUMMARY Assisted Reproduction Technology (ART) is a clinical treatment for infertile couples who want to achieve a pregnancy. In ART, embryologists fertilize eggs retrieved from the patient or a donor, culture the resulting embryos in vitro, and then transfer the selected embryo(s) to the mother's uterus. While ART is responsible for 1.9% of babies born in the United States as of 2018, selecting which embryo to transfer is a significant challenge. The difficulty comes from the complexity of confounding factors and the lack of understanding of human pre-implantation embryo development. Because of this difficulty, multiple embryos are often transferred to increases the potential of success, resulting in multiple pregnancy rates of nearly 20%, which can lead to significant morbidity and medical expenses to patients. The ideal is to transfer only a single embryo, but this necessitates the ability to select the best embryo from a cohort. Here, we propose to create a clinical decision support system to improve embryo selection in ART. To this end, we will develop novel deep learning models for robust embryo feature extraction and interactive data visualization methods for human-in-the-loop analysis. We will first extract and analyze visual features from routinely collected images of embryos. We will then combine these visual features with patients' electronic health record (EHR) data to develop interpretable computation models that score embryos on their viability. We plan to integrate our machine learning solutions into an easily accessible cloud service platform that will be adaptable across clinics to improve ART embryo selection and clinical data analysis. Our research goals will be achieved by novel machine learning-based models for morphological feature extrac- tion and importance estimation of each confounding factor and a clinical decision support system for ART. For morphological feature extraction, we plan to conduct semi-supervised learning of convolutional neural networks to minimize manual labeling that requires extensive human effort. Our feature extraction model will be the first comprehensive classification and segmentation method for ART. To aid in embryo selection, we will develop novel deep learning-based models to predict probabilities of achieving pregnancy by accepting visual features and EHR data as the input. We will also develop visual analytic tools that allow analysts to better understand and steer these deep learning models. We will estimate the importance of each input interpretable factor in embryo selection to explain the prediction to embryologists. Finally, we will develop EmbryoProfiler, a clinical decision support system for ART, that combines our machine learning-based models with a user-facing suite of visual analytic tools to support user guidance and clinical decision making. EmbryoProfiler will help facilitate daily operation in clinics, foster human-guided decision making, enrich data-driven embryo analysis, and enhance the ability to select the developmentally most competent embryo for transfer to improve ART success rates. Our project will create state-of-the-art analysis approaches for ART clinicians.

项目摘要 辅助生殖技术（ART）是一种针对不孕夫妇的临床治疗方法， 怀孕在ART中，胚胎学家从患者或捐赠者那里取出卵子， 然后将选定的胚胎转移到母亲的子宫。而抗逆转录病毒疗法导致1.9%的婴儿 截至2018年出生在美国，选择移植哪个胚胎是一个重大挑战。困难 来自于混杂因素的复杂性和对人类胚胎着床前发育的认识不足 胚胎发育。由于这种困难，多个胚胎往往被转移，以增加潜在的 成功，导致近20%的多胎妊娠率，这可能导致严重的发病率和医疗费用。 患者的费用。理想的情况是只移植一个胚胎，但这需要能够选择合适的胚胎。 最好的胚胎在这里，我们建议创建一个临床决策支持系统，以改善胚胎 选择艺术 为此，我们将开发新的深度学习模型，用于稳健的胚胎特征提取和交互式 用于人在回路分析的数据可视化方法。我们将首先提取和分析视觉特征， 定期收集胚胎的图像。然后，我们将联合收割机将这些视觉特征与患者的电子健康相结合 记录（EHR）数据，以开发可解释的计算模型，对胚胎的生存能力进行评分。我们计划 将我们的机器学习解决方案集成到易于访问的云服务平台中， 跨诊所改进ART胚胎选择和临床数据分析。 我们的研究目标将通过新颖的基于机器学习的形态特征提取模型来实现， 每个混杂因素的作用和重要性估计以及ART的临床决策支持系统。 形态特征提取，我们计划进行卷积神经网络的半监督学习 以最大限度地减少需要大量人力的手动标记。我们的特征提取模型将是第一个 ART的综合分类和分割方法。为了帮助胚胎选择，我们将开发 新的基于深度学习的模型，通过接受视觉特征来预测实现怀孕的概率， EHR数据作为输入。我们还将开发可视化分析工具，使分析人员能够更好地理解和引导 这些深度学习模型。我们将估计胚胎选择中每个输入可解释因素的重要性 向胚胎学家解释这个预测最后，我们将开发EmbryoProfiler，一个临床决策支持系统， ART系统，将我们基于机器学习的模型与面向用户的可视化分析套件相结合 支持用户指导和临床决策的工具。EmbryoPro Filler将有助于简化日常操作， 诊所，促进人类指导的决策，丰富数据驱动的胚胎分析，并提高能力， 选择发育能力最强的胚胎进行移植，以提高ART成功率。我们的项目将 为ART临床医生创建最先进的分析方法。