B0367
Title: Exploring latent spaces: manipulating medical data through image editing
Authors: Tobias Weber - LMU Munich (Germany) [presenting]
Abstract: Recent advances in image editing and manipulation techniques have opened up new possibilities in various domains, including medical imaging. Manipulating latent representations of medical data, e.g. through gradient-guided walks, can isolate pathology features or visualize disease progression. Various methods are showcased, encompassing a range of generative architectures: (1) A multi-task variational autoencoder with a survival objective is utilized to visualize hazard factors in CTs with liver metastases via transforming the latent distribution. (2) Utilizing inversion of generative adversarial networks retrieves an implicit embedding of data samples. Guided image manipulation is subsequently used to manifest degrees of pathologies on chest X-rays. (3) In latent diffusion models for chest X-ray synthesis, the spatially aware embedding serves as a measure for image in- and outpainting, enabling e.g. removal of distracting medical devices. These showcases highlight the potential of image editing in medical imaging, offering valuable insights into pathology features or disease visualization and paving the way for enhanced diagnostic and interpretative capabilities.
