Unsupervised Clustering of MADLI Medical Images Using Deep Learning

Matrix-Assisted Laser Desorption/Ionization Mass Spectrometry Imaging (MALDI-MSI) presents a powerful yet complex method for visualizing molecular distributions within biological tissues. Despite its potential, the in- herent noise and high dimensionality of MALDI-MSI data pose significant challenges for effective clustering and analysis. This thesis addresses these challenges by developing a robust methodology for the unsupervised cluster- ing of MALDI-MSI data, Clustering these images is essential for uncovering meaningful biological patterns and enhancing diagnostic capabilities. However, the high noise levels and complexity of MALDI-MSI data present significant hurdles. We examine the structure of the noise-robust deep clustering network and introduce a novel approach that in- tegrates its denoising step with feature extraction using the Xception network. Additionally, we investigate the impact of various preprocessing techniques on the clustering results of MALDI-MSI data. This comparative anal- ysis includes smoothing, baseline correction, normalization, and peak picking, revealing that each method offers varying sensitivity to noise and produces different clustering results. These variations in preprocessing stages are crucial for subsequent studies, as they affect the reliability and interpretability of the clustering outcomes. The datasets used in this study, which include mouse urinary bladder and human lung tissue samples, provide a diverse range of biological contexts for validating the proposed methods. The evaluation metrics used include the Fraction of Isotopes Correctly Clustered (F) and the Relative Isotope Ra- tio (R). The F metric measures the proportion of isotopes that are accurately grouped into their respective clusters, providing an indication of clustering precision. The R metric assesses the ratio of isotopes within a cluster relative to their expected ratio, offering insight into the clustering consistency and accuracy. Our findings demonstrate that the novel denoise-Xception method achieves clustering performance comparable to state-of-the-art methods. Specifically, our approach shows results similar to the noise-robust deep clustering network in the mouse dataset and, in some cases, performs even better in the human lung tissue dataset. This highlights the enhanced capacity of the Xception-based approach and its potential for more effective clustering. These results underscore the importance of effective noise management in MALDI-MSI data, highlighting that the denoising step is crucial for more efficient clustering. Addressing the challenge of noise in MALDI-MSI data is key to enhancing clustering efficiency and reliability. By improving the denoising step, the proposed method enables more accurate feature extraction and clustering, thus providing deeper insights into the molecular com- position of biological tissues. This research contributes to the advancement of molecular imaging and diagnostics by offering a framework that improves the reliability and interpretability of MALDI-MSI data, ultimately aiding in the identification of disease mechanisms and potential biomarkers.