Topology-Aware Reconstruction of Curvilinear Structures
Dr. Doruk Öner,EPFL
Abstract: Curvilinear structures, such as roads in remote sensing images, blood vessels in medical scans, and neural pathways in microscopy images, are vital for applications including city planning, surgical planning, diagnosis, and treatment planning. Due to their complexity and vast nature, manual reconstruction is impractical, making automatic reconstruction a significant task of computer vision. However, traditional deep learning methods often fail to preserve the topological integrity of these structures, focusing instead on per-pixel accuracy. This talk will highlight the need to move beyond conventional deep learning paradigms to achieve accurate and reliable reconstructions that maintain topological features. The focus will be on topology-aware deep learning methods that enhance the delineation of curvilinear structures in various imaging modalities. I will first introduce a novel connectivity-oriented method, designed to improve network-like structure reconstruction by preventing unwanted disconnections in 2D images. The discussion will then cover the use of Persistent Homology to capture and correct topological errors with a new filtration technique that integrates spatial information. Additionally, an active contour model-based loss function will be described, which adapts imprecise annotations during training of deep networks while maintaining topological integrity. Finally, I will discuss potential future directions to extend and refine these methods for new challenges and applications. Altogether, my work demonstrates that optimizing the topology of reconstructions is not only feasible but essential for developing next-generation methods that can drive new scientific and practical insights across multiple disciplines.
Biography: Doruk Öner is a postdoctoral researcher in Computer Vision Lab at EPFL, specializing in advanced computer vision and image analysis methods. His research focuses on the topology-aware delineation of curvilinear structures, which has important applications in city planning, medical diagnosis, and treatment planning. Doruk has developed innovative methods to enhance topological accuracy in delineation and has expanded his work to include uncertainty estimation, shape generation, and aerodynamic optimization. His methods are widely used by biomedical researchers, environmental researchers, and civil engineers to advance their work. His research resulted in publications in prestigious journals and conferences such as IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Medical Imaging, ICML, and MICCAI. Previously, Doruk earned his PhD from EPFL in 2023, working under the supervision of Prof. Pascal Fua in Computer Vision Lab. His doctoral research introduced groundbreaking approaches to automated reconstructions. Before his PhD, Doruk graduated with high honors from Bilkent University in 2018, earning a Bachelor’s degree in Electrical and Electronics Engineering and ranking 10th in his cohort.
DATE: June 24, Monday @ 13:30
Place: zoom
This is an online seminar. To obtain event details please send a message to department.