Regional manifold learning for medical image analysis
Dong Hye Ye · Scholarly Commons (University of Pennsylvania) · 2013
In this thesis, we present image-based morphological analysis methods for diagnosis of diseases. Recent advances in imaging scanners enabled us to acquire high-resolution images that capture morphological characteristics of individual anatomies. However, interpreting these images is challenging due to subtlety of morphological changes and inevitable inter-rater variability. This has created the need for sophisticated and automated image analysis methods, which can precisely quantify subtle and complex patterns of structural changes. To meet this need, we propose morphological image analysis methods based on manifold learning techniques with a goal of identifying a disease. Manifold learning techniques have recently gained popularity for medical image analysis because it provides an accurate mechanism for describing the anatomies from images through a low-dimensional embedding. However, current manifold learning techniques faces several challenges in application to medical images because they measure the anatomy in the entire image domain. Then, they do not distinguish regional impacts due to disease but detecting these regional impacts is often essential in advancing disease detection. In addition, the resulting manifold for the whole anatomy is most likely an inaccurate approximation of the true data structure due to the limited sample size in medical image applications. The major contribution in this regard is the development of a way of learning the manifold specific to individual regions and fusing their outcomes for morphological analysis. We refer this new manifold learning strategy as regional manifold learning. This regional manifold learning allows us to accurately capture the morphometry in localized regions and advance the diagnosis with fused information from multiple regions. In this thesis, we outline our morphological analysis framework, discuss the computational methodologies based on regional manifold learning, and apply this framework to identify the long term effects of reconstructive surgery of Tetralogy of Fallot from cardiac MRI. Our experimental results indicate that regional manifold learning provides a powerful mechanism for identifying disease characteristics from medical images.