Finding Patterns In Mass Spectrometry Images

Abstract

Analysis of mass spectrometry (MS) images could have many applications including aiding a pathologist in diagnosis of tissue samples, guiding a surgeon during tumor removal procedures, and discovering biomarkers of disease. MS is an analytical chemistry technique that can provide insight into metabolite patterns of tissue. MS imaging (MSI) produces mappings of MS information across a tissue sample. The processing and visualization of MS images is challenging due to their high dimensionality. In this work we used MS images obtained from breast conserving surgery (BCS), a common treatment for breast cancer. Margin analysis of removed tissue from BCS has been associated with factors such as decrease in local recurrence and is commonly performed post-operatively by a pathologist. Pathologic diagnosis has reported inter-observer and intra-observer variability. The lack of high quality intraoperative analysis has led to high rates of reexcision. Analysis of tissue samples could enhance efficacy of surgical procedures such as BCS. We hypothesize that application of proposed computational methods will enable multivariate visualization of MS images with strong correspondence to gold standard annotated histology images. For this dissertation we had access to 9 tissue samples obtained from BCS with a diagnosis of invasive ductal carcinoma. We proposed a series of computations including dimensionality reduction and graph clustering for automatic foreground segmentation and foreground clustering; affine spaces, or flats, to represent metabolite patterns of tissue types from across samples; and distance maps for novel multivariate visualization of MS images. We compared our results to conventional forms of visualization and to conventional processing methods. Our computations achieved untargeted multivariate processing and visualization of MS images. Distance maps showed strong correspondence to annotated histology images and were superior to conventional visualization techniques. The affine spaces identified biologically relevant ions that could be associated with fatty acids, which are precursors to energy cycles that are enhanced in malignant cells. Our proposed computations were statistically significantly different than conventional methods. The proposed computations could be applied to BCS to provide surgical guidance and be used in conjunction with gold standard pathology analysis to increase speed of diagnosis.