Simon Gutwein
About Me
I am a PhD student specializing in computational biology and deep learning applications, with a focus on analyzing high-dimensional imaging data from advanced techniques like Imaging Mass Cytometry (IMC) and Fluorescence in Situ Hybridization (FISH). My work aims to address the critical challenges in biological image analysis by developing scalable, interpretable, and efficient computational frameworks. I am passionate about using machine learning to unlock new insights into cell-type heterogeneity, spatial organization, and genetic aberrations in complex tissues.
Research Focus
My research is centered around improving the analysis pipelines for highly multiplexed imaging data, such as IMC and FISH. Current methods often rely on segmentation-based approaches that can introduce errors and artifacts, especially when dealing with coexpression of mutually exclusive markers. My work focuses on creating innovative deep learning models that bypass the need for segmentation and provide interpretable results, offering a more reliable understanding of tissue architecture and cell types.
Selected Projects
Segmentation-Free Analysis of Imaging Mass Cytometry Data
One of my primary research projects involves developing a deep learning framework that eliminates the need for cell segmentation in the analysis of Imaging Mass Cytometry (IMC) data. Using grouped convolutions, this model independently learns the most relevant features from each imaging channel, allowing for a direct interpretation of how different proteins contribute to the cell’s representation. This approach was validated on a dataset of over 1.8 million cells from neuroblastoma patients, accurately identifying known cell types without the need for manual feature selection. It offers a scalable, interpretable, and segmentation-free method for analyzing high-dimensional imaging data.
Link to PDF: https://arxiv.org/abs/2411.03341
FISHing in Uncertainty: Synthetic Contrastive Learning for Genetic Aberration Detection
In my recent work presented at the UNSURE Workshop, I introduced a novel approach that leverages synthetic data and contrastive learning to classify genetic aberrations in FISH images. This method removes the need for manual annotations and integrates uncertainty estimation, providing a more accurate and robust classification model. The approach not only improves diagnostic workflows but also enhances model interpretability by aligning uncertainty with expert-level judgments. This work showcases the potential of synthetic data and contrastive learning in advancing the field of medical image analysis.
Link to PDF: https://arxiv.org/abs/2411.01025
GENUINE: Genetic Aberration Classification in FISH Images
In another project, I developed GENUINE, a two-stream deep learning architecture for classifying genetic alterations in Fluorescence in Situ Hybridization (FISH) images. This model addresses signal variability and the challenges posed by manual annotations, effectively incorporating uncertainty estimation into its predictions. GENUINE’s ability to generalize across unseen data and handle noisy labels makes it a powerful tool for cancer diagnostics.
You can find more details and the code for these projects on my GitHub: github.com/simonbon