Chao Chen (09/13/23): Topological Uncertainty and Representations for Biomedical Image Analysis
Accurate delineation of fine-scale structures from images is a very important yet challenging problem. Existing methods use topological information as an additional training loss, but are ultimately making pixel-wise predictions. In this talk, we present the idea of making inference with regard to structures. Under-the-hood is the usage of discrete Morse theory to decompose an input image into structural hypotheses. This allows us to learn representations at structural level, and learn uncertainties of deep neural networks with regard to these structures. Our method makes structural-level inference rather than pixel-maps, leading to better topological integrity in automatic segmentation tasks. It also facilitates semi-automatic interactive annotation/proofreading via structure-aware uncertainty.