Fast and Robust Deep Learning for Medical imaging: Segmentation and Registration methods invariant to contrast and resolution

点击翻译按钮获取中文摘要

Project Summary Title Fast and Robust Deep Learning for Medical imaging: Segmentation and Registration methods invariant to con- trast and resolution. Summary Segmentation and registration are critical tasks in a broad range of scientific studies, and have been widely implemented in imaging analysis frameworks. Unfortunately, most existing tools suffer from two important draw- backs: they are computationally demanding, and most often impose limiting restrictions on the type of image data that can be accurately analyzed. While the former drawback has been recently addressed through the use of deep neural networks that execute rapidly once trained, these systems amplify the latter, which remains a major restriction. This typically means that tools only yield accurate results on a very limited range of scan types, most commonly those that they were trained on and are susceptible to repeating bias present in those data. For segmentation this is particularly burdensome as training frequently requires manually labeled representations for different types of input data. The constraint of image type greatly restricts image analysis and its downstream impact in an array of important domains. For example, in research imaging, it limits multi-site and longitudinal studies that must hold acquisition protocols constant or attempt to harmonize protocols across different acquisition platforms, and even this process has limited success when the differences are too extreme (e.g. across field strength). Investigators often need to adjust, redesign, or retrain the tools for their intended tasks and available images and manual labels, leading to more barriers to analysis. There is also a wealth of knowledge to be gained from analyzing clinically-sourced MR images, which could lead to better understanding of the biological underpinnings of many disease processes and a more precise quantification of the efficacy of therapeutic interventions. However, scans acquired as part of routine clinical care are often of diverse contrast, significantly lower resolution, and lower quality due to noise or subject motion. There are few if any publicly available tools that can handle the wide range of acquisition variability in typical clinical imaging. We propose to design and distribute machine learning based tools to completely remove these barriers. We will develop imaging segmentation and registration deep learning methods that retain their accuracy given unpro- cessed scans of most contrasts or resolution without the need for training data or network fine-tuning to each data variation. We will build on our recent work in learning-based methods for segmentation, registration, syn- thesis, and augmentation to leverage the speed of neural networks, the richness of MR physics models, and the generalizability of probabilistic Bayesian models. We will validate these tools on a large comprehensive multi-site study incorporating new manual labeling of scans spanning different age, sex, and race . Finally, we will deploy them to analyze anatomy and white matter lesions in a retrospective stroke cohort. The novel techniques will be implemented both as standalone open source software as well as part of FreeSurfer analysis package, making them freely available to thousands of method developers as well as science and clinical researchers.