Fast and interpretable unsupervised domain adaptation for FIB-SEM cell segmentation - Archive ouverte HAL Access content directly
Conference Papers Year : 2023

Fast and interpretable unsupervised domain adaptation for FIB-SEM cell segmentation

Abstract

Electron microscopy imaging techniques allow biologists to obtain nanoscale slice volumes of cells in which biological structures can be statistically analyzed under the condition that the tedious task of segmentation can be automated. If a learned segmentation has good results on images obtained under the same acquisition process, variations in this acquisition can drastically affect the final image’s features and can lead to a drop of segmentation performances. To tackle this problem, we propose an Unsupervised Domain Adaptation (UDA) framework based on Batch Normalization. Our method tested on mitochondria segmentation shows better results than state-of-the art methods and uses few computation resources, allowing a near real-time use. Furthermore, we propose a metric based on the Wasserstein distance to evaluate the effect of this normalization. We empirically show that this measure is log-linearly correlated with the drop of performances of the adapted segmentation and sheds light on the gain obtained by our UDA framework.
Fichier principal
Vignette du fichier
ISBI_paper_559.pdf (1.48 Mo) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04017099 , version 1 (08-07-2024)

Identifiers

Cite

Alexandre Stenger, Luc Vedrenne, Patrick Schultz, Sylvain Faisan, Étienne Baudrier, et al.. Fast and interpretable unsupervised domain adaptation for FIB-SEM cell segmentation. International Symposium on Biomedical Imaging (ISBI), Carthagena, Colombia, april 2023, Apr 2023, Cartagena, Colombia. ⟨10.1109/ISBI53787.2023.10230327⟩. ⟨hal-04017099⟩
79 View
0 Download

Altmetric

Share

Gmail Mastodon Facebook X LinkedIn More