TY - THES ID - 139039262 TI - Improved classification of partially labeled data in Imaging Mass Spectrometry through integration of morphological information AU - Tiest, Wim AU - De Moor, Bart AU - Waelkens, Etienne AU - KU Leuven. Faculteit Ingenieurswetenschappen. Opleiding Master of Artificial Intelligence (Leuven) PY - 2019 PB - Leuven KU Leuven. Faculteit Ingenieurswetenschappen DB - UniCat UR - https://www.unicat.be/uniCat?func=search&query=sysid:139039262 AB - Images revealed by a range of staining techniques remain at the base of histopathology practice, and form the key input for expert pathologists to define regions of interest on samples or reach diagnostic conclusions about abnormalities of the specimen or on the nature of the disease. Analysis of the rich dataset behind mass spectrometry imaging (MSI) of the same specimens does not present with a similar convenient visual guidance. This research explores the hypothesis that information extracted from stained images can facilitate the interpretation of MSI data on the same sample by providing clues that will drive the subsequent analysis onto the optimal choice of sample regions or selection of mass/charge (m/z) values. To this purpose, stained images and MSI data are processed in a parallel feature extraction and labeling pipeline, and the output of those pipelines is driving an ultimate joint analysis based on clustering techniques. This report shows that texture carries the more relevant image information. The extraction of texture information using a Haralick transformation variant, and subsequent segmentation based on a spatially aware waterflood algorithm delivers an unsupervised labeling of the image. Adding the original density information into the pipeline improves the detection of structural features. The vast pool of information inside a full MSI dataset can be summarized with a digitization method based on both spatial and spectral dimensions, a fingerprint extracting only high relevant peaks per pixel and a Term Frequency / Inverse Document Frequency (TF-IDF) and k-means clustering and labeling algorithm borrowed from the field of text analysis. A visualization of the full MSI data is obtained much faster than published t-SNE or UMAP techniques with full retention of the link to real mass values. Integrating the features of both initial pipelines using another variant of the TF-IDF and k-means clustering methodology confirms the research hypothesis on 1 of the 2 provided test samples. Alignment of the source features at pixel level seems to be a main reason for not achieving proof of concept on the second sample. Both the intermediate image and MSI labeling, the image alignment methodology and the final clustering of the successful sample present attractive options for further analysis and follow-up research. ER -