Book Description
"Soft-tissue sarcoma is a rare cancer that has high metastatic potential to the lung with poor prognosis at 3-year survival for patients that develop lung metastasis. As certain prognostic factors such as necrosis and poor perfusion due to abnormal blood supply may be targets for novel strategies, research in imaging techniques that can characterize the tumour microenvironment can allow for prediction of patient response to treatment. Dynamic Contrast-Enhanced (DCE) MRI is a functional imaging technique that can visualize perfusion in biological tissue by acquiring multiple images following the injection of a contrast agent through the cardiovascular system to analyze the contrast uptake within a given tissue. DCE-MRI has a complex 4-dimensional dataset consisting of thousands of pixels per image across multiple timepoints. DCE-MRI analysis tends to aim towards building semi-quantitative methods and model-based approaches to describe the kinetics of the contrast agent. However, the implementation of these techniques often requires a-priori information that puts constraints on the data. In this thesis, a data-driven technique to DCE-MRI analysis is proposed using non-negative matrix factorization (NMF), a dimensionality reduction technique that can isolate signal patterns in the data and provide visualization of the spatial distribution of these patterns. Using a DCE-MRI dataset consisting of sarcoma tumours over the course of radiotherapy, we show that the alternating non-negative least squares using block pivot principle (ANLS-BPP) NMF framework can find high and low signal enhancement curves in this data and generate weight maps for each perfusion curve that can be superimposed to visualize the heterogenous spatial distribution of high and low perfusion in these tumours. While these signal enhancement time-course patterns are consistent across patients and over the course of the radiotherapy, the weight maps across several timepoints carry the changes in perfusion distribution in response to radiotherapy over the course of treatment. However, these weight maps vary according to the random initialization of the NMF algorithm due to the non-uniqueness of the solutions to the algorithm as it tends to converge onto local minima. For this reason, we proposed a multi-NMF algorithm that performs an averaging of the weight maps produced by the algorithm using a distance minimization function with multiple tolerances to obtain the most representative weight map for each sarcoma tumour. This algorithm could reduce the variability in the weight maps produced by the NMF algorithm, thereby increasing the robustness of this technique to produce repeatable perfusion distributions in sarcoma tumours. These results have significant implications in the development of model-free approaches to DCE-MRI analysis as the ANLS-BPP NMF algorithm can find consistent signal enhancement patterns in the data and generate weight maps that can spatially visualize the perfusion distributions. Furthermore, the proposed multi-NMF algorithm can, in principle, be applied to any NMF algorithm to reduce the variability of the perfusion maps. Future work aims to test the ANLS-BPP algorithm on different types of solid tumours and investigate the ability for the proposed multi-NMF algorithm to improve the variation issues on other frameworks of the NMF algorithm"--