JULAIN Talk by Suprosanna Shit
Relationformer: A Unified Framework for Image-to-Graph Generation
Suprosanna Shit, Image-Based Biomedical Modeling Group (IBBM), TU Munich
- When: Oct 17, 2020, 10am
- Where: INM Seminarroom building 15.9, room 4001b
Abstract
Segmenting tubular structure is a recurring problem in medicine, biology, and remote sensing. Image segmentation serves as an intermediate representation of the task at hand, which broadly deals with the underlying network structure or structural graph. For example, vessel networks for medicine, and neuronal networks in biology, heavily rely on segmentation, and correct network topology is often of paramount interest. However, the traditional loss function gives equal weightage to false positives and false negatives pixels, irrespective of its importance in preserving the topological structure during segmentation. Finding an efficient loss to enforce topological correctness is the central theme of this talk. In an alternative direction, one can ask the following research question: Can we directly infer the underlying graph representation without an explicit segmentation stage? If so, how far can we go? Notably, graph extraction from an image is a relatively new field of research and mainly involves semantic knowledge graphs, i.e., scene-graph extraction from natural images. Can we merge these two parallel endeavors under a single framework? Moreover, can we benefit from translating image-to-graph models from the scene-graph community to structural graph extraction? This question will lead to the second part of the talk.
Papers
- Suprosanna Shit, Rajat Koner, Bastian Wittmann, Johannes Paetzold, Ivan Ezhov, Hongwei Li, Jiazhen Pan, Sahand Sharifzadeh, Georgios Kaissis, Volker Tresp, Bjoern Menze. Relationformer: A Unified Framework for Image-to-Graph Generation paper
- clDice - a Novel Topology-Preserving Loss Function for Tubular Structure Segmentation
Invitation and moderation: Hanno Scharr