Description |
-
Abstract: Medical images are collected via a broad assortment of modalities with the objective of studying structures ranging from cells to tissues, to entire organs for applications as diverse as elucidating disease mechanisms, understanding the effectiveness of treatment options, as well as drug discovery. In these applications, clinical relevant information needs to be extract from large amount... read mores of data for statistical analysis. However, approaches based on human-only interaction with the data suffer from a number of problems including the quantity of effort and time required to process the large data sets as well as the the related problem of inconsistency of the analysis due to human error. The use of well designed computer-based methods for processing these data addresses both of these issues. First, with the developments of hardware and computational techniques, more sophisticated and powerful algorithms can be applied to larger data sets. Second, results based on algorithms are always reproducible given the specification of parameters defining the operation of these methods. Although humans are still needed for works such as algorithm selection and parameter tuning, repetitive manual effort can be significantly reduced. In this thesis, we present algorithms using level set methods for two representative applications of medical image analysis: analysis of magnetic resonance (MR) images of the human knee to quantify osteoarthritis and analysis of phase contrast microscopy (PCM) neuron images for Ivermectin effect evaluation. For the MR knee images, we propose a new coupled prior shape model which incorporates prior shape information and also the relative position information for multiple objects. In addition, our new shape model implicitly puts on more constraints in areas with less shape differences while conventional shape models always impose constraints uniformly on the whole shape domain. Moreover, an edge based force incorporating directional information is also introduced. Segmentation results on the real MR knee images demonstrate the feasibility of our new coupled prior shape model and also the directional edge-based force. Based on the combination of our coupled prior shape model and the directional edge-based force, we can reduce the manual interaction time significantly for bone marrow lesions segmentation without sacrificing accuracy. For the PCM neuron images, we introduce a parametric image model using the level set framework to represent images to be restored and segmented. Moreover, we formulate an optimization problem which merges image restoration and segmentation and give its solutions. Based on the above segmentation method, we propose a pipeline which is automatic and the first to segment somas and trace dendrites simultaneously for the PCM neuron images. Results for both the synthetic and real images validate and demonstrate the advantages of our approach.
Thesis (Ph.D.)--Tufts University, 2015.
Submitted to the Dept. of Electrical Engineering.
Advisor: Eric Miller.
Keywords: Electrical engineering, Biomedical engineering, and Medical imaging and radiology.read less
|
This object is in collection