Vector-Quantized Latent Flows for Medical Image Synthesis and Out-Of-Distribution Detection

Firas Khader, Gustav Müller-Franzes, Soroosh Tayebi Arasteh, Tianyu Han, Jakob Nikolas Kather, Johannes Stegmaier, Sven Nebelung, Daniel Truhn

Abstract

We present an innovative method that allows for simultaneous out-of-distribution detection and image generation by encoding images in the latent space of a vector-quantized autoencoder and using normalizing flow models. The technique is demonstrated on a medical dataset of knee radiographs and can be used to relieve clinical radiologists of tedious tasks of quality control while simultaneously guiding radiologic technologists to improved and standardized image quality during image acquisition.
BibTeX
@inproceedings{khader2023vector,
  author = {Firas Khader and Gustav Müller-Franzes and Soroosh Tayebi Arasteh and Tianyu Han and Jakob Nikolas Kather and Johannes Stegmaier and Sven Nebelung and Daniel Truhn},
  title = {{Vector-Quantized Latent Flows for Medical Image Synthesis and Out-Of-Distribution Detection}},
  booktitle = {ISBI},
  year = {2023},
  doi = {10.1109/isbi53787.2023.10230460},
}

Loading related content...

Figure