SCH: EAGER: RUI: Collaborative Research: A novel 3D image predictive model for osteoarthritis disease

膝盖骨关节炎（OA）影响了10%的老年人，是导致缺勤、提前退休和关节置换的主要原因。膝关节骨性关节炎是一种以膝关节软骨退化为特征的疾病。使用目前的技术，很难预测会发生多快或多大程度的恶化，因为软骨损失是一个缓慢而渐进的过程，只能通过医学图像来检测。本项目将探索一种新的3D图像模型，可以准确预测膝关节软骨的变化，促进OA的早期发现和个性化治疗。如果成功，该项目可以通过降低OA治疗相关的高经济成本，并改善这些人的生活质量，使美国3500万人受益。这些学院计划将研究成果传播给当地医疗界，并设计一门新的课程，让本科生参与研究。新型的三维图像预测模型应具有广泛的成像应用。本项目旨在探索一种新的医学成像三维信息融合机制和一种以深度神经网络为核心的三维图像到图像预测模型。该项目将整合来自MRI序列的软骨信息。为了处理尺寸差异和进行图像配准，将定义一个通用坐标系，以形成软骨平面的连续完整表示。使用坐标系统，深度神经网络将被训练来学习三维软骨图之间的潜在相关性。与传统的图像到单值预测不同，该模型将进行图像到图像的预测；即从当前的3D软骨图到未来的3D软骨图，不同的时间长度（分别为2年、4年、6年和8年），利用大型成像数据库。最后，该团队将从软骨图构建未来的3D膝关节模型，以3D视图显示软骨变化的轨迹。

Knee osteoarthritis (OA) affects 10% of older adults and is a major cause of work absence, early retirement and joint replacement. Knee OA is a disease characterized by deterioration of the cartilage in the knee. Using current technology, it is hard to predict how fast or how much deterioration will take place because cartilage loss is a slow and gradual process and can only be detected through medical images. This project will explore a novel 3D image model that can predict the accurate change of knee cartilage, to facilitate early detection and personal treatment for OA. If successful, the project could benefit 35 million people in the United States by reducing the high economic cost related to OA treatment, and improving the quality of life for these people. The PIs plan to disseminate the research to local medical communities and design a new course to involve undergraduate students into the research. The novel 3D image predictive model should have a wide variety of imaging applications. The goal of this project is to explore a novel 3D-information-fusion mechanism for medical imaging and a novel 3D image-to-image predictive model using deep neural networks as the core. The project will integrate cartilage information from MRI sequences. To handle size differences and perform image registration, a universal coordinate system will be defined to form a continuous and complete representation of the cartilage plane. Using the coordinate system, deep neural networks will be trained to learn the underlying correlation between the 3D cartilage maps. Unlike the traditional image-to-single-value prediction, the model will make image-to-image prediction; that is, from a current 3D cartilage map to a future 3D cartilage map, for different lengths of time (2, 4, 6, and 8 years respectively), leveraging a large imaging database. Finally, the team will construct the future 3D knee models from the cartilage maps to display the trajectory of cartilage change in a 3D view.