Machine learning-based segmentation and risk modeling for real-time prediction of major arterial bleeding after pelvic fractures

项目总结/摘要：骨盆骨折后动脉出血是导致 钝伤致死动脉出血风险的预测是困难的，目前使用 主观标准，通常基于入院计算机断层扫描（CT）的定性结果。分段 血肿和造影剂外渗（CE）量预测是否需要血管栓塞、大量输血， 死亡率，但不能实时应用。盆腔血肿的多灶性不明确与CE 妨碍了使用基于直径的测量的可靠估计。Dreizin博士是一名创伤放射科医生， 马里兰州大学医学院。他的早期工作集中在提高速度和可靠性， 使用半自动化技术对盆腔血肿进行体积分析，并推导出Logistic回归方程。 骨盆骨折后主要动脉损伤的回归预测工具。德雷辛博士的目标是这四个- 2008年K 08辅导职业发展奖的建议是获得所需的技能，1）实施深入 用于自动血肿体积分割的学习架构，以及2）开发计算模型 用于骨盆创伤后的预后预测。这些工具可以大大提高速度和准确性 在危及生命的创伤性盆腔出血情况下的临床决策。全卷积神经 网络（FCN）已经成为自动化医学图像的最鲁棒和可扩展的方法 细分直观的软件平台，用于培训FCN实施和生成多变量 机器学习模型是在Python编程环境中开发的。培训 本提案的目标和研究活动是必要的，以提供新的技能， Python编程、深度学习软件和计算建模软件的实践经验。通过 了解现代机器学习背后的原理和计算基础设施，Dreizin博士将 能够独立培训和验证最先进的算法，并有效地领导团队， 这方面的研究人员。为了实现他的目标，Dreizin博士组建了一个多学科的导师团队， 顾问和合作者，在医学成像，概率，计算机视觉领域拥有世界领先的专业知识 理论，数据科学和比较有效性研究。德雷辛博士将重点关注两个具体目标。在 目标1，他将训练和验证用于创伤性盆腔血肿分割的深度学习架构 和CE通过计算Dice度量、时间努力和与临床结果的相关性。在目标2中，他将 基于丰富的临床试验数据，生成并测试用于预测骨盆创伤后大动脉出血的定量模型。 分割特征的多标签数据集。训练和试验数据将是必要的德雷津博士的长期- 研究独立性和R 01支持的长期目标是为 骨盆创伤后的临床重要影像学特征谱，以及全自动多变量 临床预测工具，有可能转化为工业和FDA批准的产品。

PROJECT SUMMARY/ABSTRACT: Arterial hemorrhage after pelvic fractures is a leading reversible cause of death after blunt trauma. Prediction of arterial bleeding risk is difficult, and currently determined using subjective criteria, often based on qualitative results of admission computed tomography (CT). Segmented hematoma and contrast extravasation (CE) volumes predict need for angioembolization, major transfusion, and mortality but cannot be applied in real-time. The ill-defined multi-focal nature of pelvic hematomas and CE prevents reliable estimation using diameter-based measurements. Dr. Dreizin is a trauma radiologist at the University of Maryland School of Medicine. His early work has focused on improving the speed and reliability of volumetric analysis of pelvic hematomas using semi-automated techniques, and derivation of a logistic regression-based prediction tool for major arterial injury after pelvic fractures. Dr. Dreizin’s goal for this four- year K08 mentored career development award proposal is to gain the skills needed to 1) implement deep learning architectures for automated hematoma volume segmentation and 2) develop computational models for outcome prediction after pelvic trauma. These tools could greatly improve the speed and accuracy of clinical decision making in the setting of life-threatening traumatic pelvic bleeding. Fully convolutional neural networks (FCNs) have emerged as the most robust and scalable method for automated medical image segmentation. Intuitive software platforms for training FCN implementations and generating multivariable machine learning models have been developed in the Python programming environment. The training objectives and research activities of this proposal are necessary to provide Dr. Dreizin with new skills and practical experience in Python programming, deep learning software, and computational modeling software. By understanding the principles and computational infrastructure behind modern machine learning, Dr. Dreizin will be able to train and validate state-of-the-art algorithms independently and effectively lead a team of researchers in this area. To achieve his goals, Dr. Dreizin has assembled a multidisciplinary team of mentors, advisors, and collaborators with world-leading expertise in computer vision in medical imaging, probability theory, data science, and comparative effectiveness research. Dr. Dreizin will focus on two specific aims. In Aim 1, he will train and validate deep learning architectures for segmentation of traumatic pelvic hematomas and CE by computing the Dice metric, time effort, and correlation with clinical outcomes. In Aim 2, he will generate and test quantitative models for predicting major arterial bleeding after pelvic trauma based on a rich multi-label dataset of segmented features. The training and pilot data will be necessary for Dr. Dreizin’s long- term goal of research independence and R01 support to develop automated segmentation algorithms for the spectrum of clinically important imaging features after pelvic trauma, as well as fully automated multivariable clinical prediction tools with potential for translation to industry and as an FDA-cleared product.