The paper “Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation” by Arne Schmidt, Pablo Morales-Álvarez, and Rafael Molina, was accepted for ICCV 2023 in Paris.
I am very happy to announce that our paper “Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation” was accepted for ICCV 2023!
In the medical field, experts often disagree when assessing images. As studies in histopathology have shown, even the same medical expert often comes to a different conclusion when assessing the same image twice. Therefore, we hope to make the diagnostic processes more reproducible with AI. But the missing ‘ground truth’ also imposes a major challenge when training AI models.
To tackle this challenge, we propose a Probabilistic Inter-Observer and iNtra-Observer variation NetwOrk (Pionono). Pionono, inspired by the probabilistic U-Net architecture, models the labeling behavior of each annotator by a multivariate Gaussian distribution optimized in an end-to-end training process, as shown in the figure. This allows for a robust training procedure with imperfect labels and accurate prediction of possible ambiguities.
The preprint of our article is available at: https://arxiv.org/abs/2307.11397
Thanks to my great supervisors Rafael Molina and Pablo Morales Álvarez for the great collaboration and all the successful projects in the last three years.
The proposed Pionono model. The labeling behaviour of each annotator is represented by a multivariate Gaussian distribution. The drawn samples of these distributions are concatenated with the extracted features and then fed into the segmentation head. The output simulates the inter- and intra-observer variability of annotations and is optimized using the real annotations of each rater. The model is trained end-to-end with a combination of log-likelihood loss (LL) and Kulback Leibler (KL) divergence between posterior and prior, combined in the overall ELBO loss
Arne Schmidt – ESR8