The mission of the Center for Biosystems and Biotech Data Science, which is located at the Global Campus of Ghent University in Korea, is to pursue the development of novel mathematical and computational approaches for modeling of biosystems and biotech data sets. A core technology leveraged by researchers at the center is deep machine learning, targeting the development of innovative concepts, methodologies, and tools in both the area of molecular biology and the field of computer vision. Furthermore, the Center for Biosystems and Biotech Data Science, which has a headcount of three professors and ten PhD candidates, is responsible for organizing nine courses at Ghent University Global Campus (for a total of 65 ECTS), ranging from Informatics to Bioinformatics and Probability & Statistics.
Current research lines include:
Trustworthy and interpretable AI: This line of research focuses on developing trustworthy and interpretable AI models, particularly for applications in medical and biomedical imaging. We aim to enhance explainability by making the decision-making processes of AI systems more transparent and understandable to humans. By improving how AI models communicate their reasoning, this work seeks to build confidence in their use for critical healthcare tasks such as medical diagnosis. Our goal is to enable reliable and interpretable AI-driven diagnostics and imaging techniques for both medical professionals and patients.
Characterization of sub-visible particles in biopharmaceuticals: This line of research explores the application of AI techniques in analyzing subvisible particles (SvPs) formed in protein-based therapeutics such as vaccines. Sub-visible particles can originate from various sources such as protein aggregation, silicone oil, glass particles, air bubbles, and microplastics. These particles, which can result from stress factors like heat or mechanical agitation, pose risks such as loss of drug efficacy and immune reactions. By leveraging advanced imaging techniques such as Flow Imaging Microscopy and state-of-the-art AI techniques, we aim to enhance the detection, classification, and morphological analysis of SvPs to ensure the safety and effectiveness of therapeutic proteins.
Deep learning and sensor technologies for object detection and classification: This research line focuses on applying deep learning technologies to solve detection and classification problems in real-world environments, including item size measurement in industrial settings, microplastics detection for environmental monitoring, and analysis of human stress states. The work involves optimizing and applying state-of-the-art sensor technologies and deep learning models to meet the specific requirements of each field, aiming to overcome the limitations of existing methodologies and enhance performance.
Deep learning for biological sequence analysis: This research focuses on leveraging deep machine learning techniques to solve biological problems, with a particular interest in biological sequence analysis and using data-driven approaches to understand genomic sequences and their functionalities. The work involves multitask learning models that generalize across various related tasks to improve prediction effectiveness and robustness. Another area involves generative AI, developing models capable of generating biologically valid data. Key applications include the prediction of splice sites and translation initiation sites, critical for understanding gene regulation and expression.
Improving reliability and generalization of deep neural networks: Through an integrated approach combining model-centric and data-centric strategies, this research line aims to enhance the reliability and generalization of deep neural networks in real-world applications (e.g., unstained parasite detection in microscopy video). On the model-centric side, the work investigates the limitations of current evaluation methods, develops new frameworks to improve assessment of deep neural networks, and explores uncertainty quantification techniques to better understand model predictions. Complementing this, the data-centric research critically examines dataset quality, revealing important insights such as the multi-label nature of supposedly single-label benchmarks like ImageNet, which can significantly misrepresent model performance.
Neural rendering for real-time organ reconstruction: This research line focuses on advanced 3D scene rendering and intraoperative organ reconstruction in the context of laparoscopic video analysis, in collaboration with Ghent University Hospital. Specifically, it explores state-of-the-art techniques such as Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting for rendering detailed abdominal scenes. Additionally, the work on intraoperative organ reconstruction employs diffusion models, implicit neural representations (INRs), and deep weight spaces to create highly accurate, real-time organ models during surgery. The aim is to enhance precision and outcomes in minimally invasive procedures.
Surgical video analysis: This research line focuses on applying machine learning-based computer vision techniques to various aspects of laparoscopic and chemotherapy video analysis. Key areas include developing models for accurate organ and tumor recognition, automatic tumor size estimation, and phase recognition in surgical videos. Additionally, this research line investigates self-supervised learning methods to estimate depth in laparoscopic videos, addressing the challenge of limited labeled data. Through the integration of these different advancements, we aim to enhance the precision and efficiency of surgical procedures, producing better patient outcomes.
