I am thrilled to announce that our paper “Uncertain but Useful: Leveraging CNN Training Variability into Data Augmentation“ has been accepted as a poster at the International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging (UNSURE), MICCAI 2026.

Description

This work studies numerical uncertainty during the training of FastSurfer, a CNN-based neuroimaging segmentation model. It shows that FastSurfer can exhibit greater variability than FreeSurfer in cortical regions, that inexpensive random-seed perturbations reproduce similar regional variability patterns, and that ensembles built from this variability can improve downstream brain-age regression through data augmentation.

Authors

  • Inés Gonzalez Pepe
  • Vinuyan Sivakolunthu
  • Yohan Chatelain
  • Tristan Glatard