Research Interests
- Cancer Genomics
- Deep Learning for Medical Image Analysis
- Topological Data Analysis (TDA)
Research Area
My research interests lie both in methodological and applied machine learning. In terms of methodologies, I am interested in the development of novel Graph Neural Network (GNN) models and Topological Data Analysis methods (TDA) for modeling and feature extraction, respectively. Application-wise, I focus on multimodality analysis using medical imaging (CT, PET and fMRI) and genomics data, primarily for cancer genomics studies. Currently, I am involved in the three following subprojects that independently contribute to my PhD thesis.
Quantitative analysis of pre-therapeutic computed tomography (CT) imaging using deep learning-based radiomics
The current standard of care mainly focuses on staging bronchial carcinomas from pre-therapeutic imaging using the TNM and UICC classifications to assess prognosis. It has already been shown that CT scans of patients with bronchial carcinoma contain more information than can be obtained from the human brain, as well as allow classification in terms of outcome. A conventional way of doing analysis is based on previously defined radiomics parameters. Alternatively, a Convolutional Neural Network (CNN) can be trained, which independently learns essential image parameters. The example of such a model has been trained multicentrally by Stanford Hospital, H. Lee Moffitt Cancer Center and Research Institute, MAASTRO and Charité using patients (n=709) with NSCLC who have either undergone surgery or radiotherapy. In this subproject, I focus on the extension of CNN model for the particular dataset, with the goal of testing the effectiveness of transfer learning and other types of deep learning models (e.g. CapsuleNets and Transformers) for the improved survival prediction accuracy.
Voxel-based analysis of the three-dimensional dose distribution
A voxel-based homogeneity indexing is a commonly practiced method in the analysis of radiation plans. Depending on the localization and size of the tumor, dose reductions in the target volume must sometimes be accepted in order to avoid normal tissue complications, while underdoses in the target volume are associated with reduced local control. For this reason, a compromise is made between tumor control and the expected side effects. It is currently unknown what level of underdosing in the target volume can still be accepted without significantly worsening the patient's outcome, however, using the voxel-based homogeneity index, underdoses can be quantified and analyzed in terms of oncological outcome (OS, LC, LRC , PFS). In this subproject, I combine TDA methods with homogeneity index analysis in order to study the efficiency of radiation dosing in various target volumes of a tumor.
Topological Data Analysis in fMRI data for the functional discovery of latent structures.
fMRI is the preeminent method for collecting signals from the human brain in vivo , for using these signals in the service of functional discovery, and relating these discoveries to anatomical structure. Numerous computational and mathematical techniques have been deployed to extract information from the fMRI signal. Yet, the application of Topological Data Analyzes (TDA) remain limited to certain sub-areas such as connectomics (that is, with summarized versions of fMRI data). While connectomics is a natural and important area of application of TDA, applications of TDA in the service of extracting structure from the (non-summarized) fMRI data itself are heretofore nonexistent. “Structure” within fMRI data is determined by dynamic fluctuations in spatially distributed signals over time, and TDA is well positioned to help researchers better characterize mass dynamics of the signal by rigorously capturing shape within it. The goal is to illustrate how TDA can be utilized to extract structure from fMRI data and how such extraction could play an important role in the endeavor of functional discovery.
Teaching Assistantship
Supervision
Advanced Machine Learning Seminar (Winter semester 2022/2023)
Topic : Progression-free survival analysis using CapsuleNets on NSCLC patients PET-CT scans
Students : Tobias Fiedler, Franz Sauerwald.
