Machine Learning and its Application in Automatic Change Detection in Medical Images

Change detection is a fundamental problem in various fields, such as image surveillance, remote sensing, medical imaging, etc. The challenge of change detection in medical images is to detect disease-related changes while rejecting changes caused by noise, patient position change, and imaging acquisition artifacts such as field inhomogeneity.

In this thesis, first, we overview the existing change detection methods, their underlying mathematical frameworks and limitations. Second, we present our contributions in solving the problem. We design optimal subspaces to approximate the background image in more efficient fashion. This is based on our structure principal component analysis, aiming to capture the structural similarity between scans in the context of change detection. We theoretically and numerically discuss the proper choices of norms used in the subspace approximation.

The mathematical frameworks developed in this thesis consist of: (i) a new mathematical model to change detection by defining it as an optimization problem involving a cost function, input and output image sets, projection onto a subspace, and a similarity measure; (ii) development and implementation of numerical pipelines to compute the clinical changes by designing four mathematical algorithms; (iii) refining our algorithms by introducing the co-registration step utilizing the local dictionaries; and (iv) two new structure subspace learning models that are robust to outliers and noise, reduce the dimensionality of the dataset, and computationally efficient. We defined the co-registration step as a minimization problem involving a cost function, input and output image sets, a set of transform functions, projection onto a subspace, and a similarity measure.

Based on the mathematical frameworks discussed above, numerical schemes are developed to automatically filter out clinically unrelated changes and identify true structure changes that may be of clinical importance. Our approaches are data-driven and utilize the knowledge of machine learning. We quantitatively analyze the performance of these algorithms using both synthetic and real human data. Our work has the potential to be used in computer aided diagnosis.