On Direct Distribution Matching for Adapting Segmentation Networks

Abstract

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 both the quality, stability and efficiency of training. We juxtapose our approach to state-of-the-art segmentation adaptation via adversarial training in the network-output space. In the challenging task of adapting brain segmentation across different magnetic resonance images (MRI) modalities, our approach achieves significantly better results both in terms of accuracy and stability.

includes appendix; published at MIDL2020: https://2020.midl.io/papers/pichler20.html

References in corpus (9)

• A Survey on Deep Learning in Medical Image Analysis
• Learning Transferable Features with Deep Adaptation Networks
• ADVENT: Adversarial Entropy Minimization for Domain Adaptation in Semantic Segmentation
• FCNs in the Wild: Pixel-level Adversarial and Constraint-based Adaptation
• Curriculum Domain Adaptation for Semantic Segmentation of Urban Scenes
• A DIRT-T Approach to Unsupervised Domain Adaptation
• 3D fully convolutional networks for subcortical segmentation in MRI: A large-scale study
• Task Driven Generative Modeling for Unsupervised Domain Adaptation: Application to X-ray Image Segmentation
• PnP-AdaNet: Plug-and-Play Adversarial Domain Adaptation Network with a Benchmark at Cross-modality Cardiac Segmentation