Communication Dans Un Congrès Année : 2020

On Direct Distribution Matching for Adapting Segmentation Networks

Résumé

Minimization of distribution matching losses is a principled approach to domain adaptation in the context of image classification. However, it is largely overlooked in adapting segmentation networks, which is currently dominated by adversarial models. We propose a class of loss functions, which encourage direct kernel density matching in the network-output space, up to some geometric transformations computed from unlabeled inputs. Rather than using an intermediate domain discriminator, our direct approach unifies distribution matching and segmentation in a single loss. Therefore, it simplifies segmentation adaptation by avoiding extra adversarial steps, while improving quality, stability and efficiency of training. We juxtapose our approach to state-of-theart segmentation adaptation via adversarial training in the network-output space. In the challenging task of adapting brain segmentation across different magnetic resonance imaging (MRI) modalities, our approach achieves significantly better results both in terms of accuracy and stability.

Fichier principal
Vignette du fichier
On Direct Distribution Matching for Adapting.pdf (3 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04137011 , version 1 (22-06-2023)

Licence

Identifiants

  • HAL Id : hal-04137011 , version 1

Citer

Pichler Georg, Dolz Jose, Ismail Ben Ayed, Pablo Piantanida. On Direct Distribution Matching for Adapting Segmentation Networks. Conference on Medical Imaging with Deep Learning 2020, Jul 2020, Montréal, Canada. ⟨hal-04137011⟩
59 Consultations
108 Téléchargements

Partager

  • More