Medical image analysis often requires developing elaborate algorithms that are implemented as pipelines. A growing number of large medical imaging studies necessitate development of robust and flexible pipelines. In this thesis, we present contributions of two kinds: (1) an open-source framework for building pipelines to analyze large-scale medical imaging data that addresses these challenges, and (2) two analyses of large medical image collections using our tool.
Our medical image analysis pipeline construction tool, PipeBuilder, is designed for constructing pipelines to analyze difficult data where iterative refinement and development are necessary. We provide a lightweight scripting framework that enables the use of existing and novel algorithms in pipelines. We also provide a set of visualization tools to visualize the pipeline's structure, data processing status, and intermediate and final outputs. These visualizations enable interactive analysis and quality control, facilitating computation on large collections of heterogeneous images.
We employ PipeBuilder first to conduct an analysis of white matter hyperintensity in low-resolution brain images from stroke patients. Our analysis of this cerebrovascular pathology consists of three main components: accurate registration, to enable data fusion and population analysis; segmentation, to automatically delineate pathology from the images; and analysis, to extract insight using the images and the derived products. Our analysis explores the relationship between the spatial distribution, quantity, and growth of white matter hyperintensity.
Our next application of PipeBuilder is to a neuroimaging study of Alzheimer's patients, where we explicitly characterize changes over time using longitudinal data. As with the previous application, we introduce an analysis workflow involving registration, segmentation, and analysis. Our registration pipeline aligns the large, heterogeneous group of populations while still accurately characterizing small changes in each patient over time, and our analysis exploits this fact to explore change in pathology over time.
Thesis Supervisor: Polina Golland
Thesis Committee Members: Rob Miller, Ron Kikinis