TRD 4: Platforms for multi-modal and multi-scale imaging data

点击翻译按钮获取中文摘要

TRD 4: Platforms for multi-modal and multi-scale imaging data Lead Principal investigator: Susumu Mori, Professor of Radiology; Co-Principal investigators: Brian Caffo, Professor of Biostatistics; Jeremias Sulam, Assistant professor of Biomedical Engineering Co-investigators: Andreia Faria, Michael Miller, Tilak Ratnanather, Laurent Younes The role of TRD4 is to develop new technologies and platforms to integrate and analyze complex multi-modal and multi-scale imaging data via collaboration with other TRDs and CPs. In the past two decades, we have witnessed remarkable advances in image acquisition, processing, and analysis technologies for brain MRI. As we enter a new decade, however, there remain several key areas in combining information across the macro- meso-micro scales, and discovering predictive models for which we require significant advances in existing tools and technologies. One of the significant opportunities going forward is to leverage the rapidly evolving data science technologies which are now emerging and opening new frontiers for researchers. This will require advances not only in collections of software to analyze each dimension but also a new generalized framework to integrate and explore the data. Another important development in recent years is the surge of heavily data- driven approaches such as deep learning. We believe this is a great time to invest time and resources to evaluate their capability by comparing them with conventional engineering approaches. More importantly, there is great potential in combining these two approaches to test improvements in precision, accuracy, and/or efficiency. We are uniquely positioned to take a lead in this sphere by leveraging our experiences and resources (tools and data) accumulated in the past 20 years. Our specific aims will be: (1) To integrate and test deep learning (DL) approaches in image data acquisitions and analyses to solve inverse problems for estimation of latent variables in brain MRI. These variables include enhancement of SNR, anatomical resolutions, and underlying anatomical features such as axonal alignments; (2) To develop technologies and platforms for characterizing models that predict the key factors determining brain diseases by integrating multiple imaging modalities, time-domain data, and non-imaging information through statistics. Our models will focus on etiology, pathology and prognosis considering retrospective, cross-sectional and prospective data; (3) To integrate DL approaches to brain mapping strategies such as registration, image segmentation, and lesion detections. Using our rich resources for annotated image libraries (atlases), conventional segmentation / detection tools, and expertise, we will develop DL-based approaches and evaluate their efficacy, including comparison with conventional approaches in terms of accuracy and efficiency. Improvement of the performance by combining these two approaches will be tested. Furthermore, we will develop a framework for diffeomorphic image registration that can incorporate extracted features (i.e. labels and lesions) to perform brain mapping of cases with severe pathological conditions.