Joint Appearance-Feature Domain Adaptation: Application to QSM Segmentation Transfer

Bin Xiao, Naying He, Qian Wang, Zhong Xue, Lei Chen, Fuhua Yan, Feng Shi, Dinggang Shen

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Quantitative susceptibility mapping (QSM) is a magnetic resonance imaging technique used to quantitatively measure the iron content in the brain. Patients with Parkinson’s disease are reported having increased iron deposition, especially in substantia nigra (SN) which is a relatively small gray matter structure located in the midbrain. The automatic segmentation of SN is a critical prerequisite step to facilitate the progression of evaluating the course of Parkinson’s disease. However, the imaging protocol and reconstruction methods in QSM acquisition vary largely, rendering great challenges in constructing and applying image segmentation models. Thus, a model trained on a certain dataset often performs poorly on datasets from other scanners or reconstruction methods. To quickly transfer a trained segmentation model to a dataset acquired in a new instrument, we have developed a joint appearance-feature domain adaptation framework (JAFDAF) to transfer the knowledge from the source to the target domains for improved SN segmentation. In particular, we perform domain adaption in both appearance and feature spaces. In the appearance space, we use region-based histogram matching and a neural network to align the grayscale ranges of images between these two domains. In the feature space, we propose a domain regularization layer (DRL) by utilizing the idea of neural architecture search (NAS) to enforce the convolution kernels for learning features that are efficacious in both domains. Ablation experiments have been carried out to evaluate the proposed JAFDAF framework, and the experimental results on 27 subjects show that our method achieves up to 12% over the baseline model and about 5% over a fine-tuning approach.

Original languageEnglish
Title of host publicationMachine Learning in Medical Imaging - 11th International Workshop, MLMI 2020, Held in Conjunction with MICCAI 2020, Proceedings
EditorsMingxia Liu, Chunfeng Lian, Pingkun Yan, Xiaohuan Cao
PublisherSpringer Science and Business Media Deutschland GmbH
Pages241-249
Number of pages9
ISBN (Print)9783030598600
DOIs
Publication statusPublished - 2020
Externally publishedYes
Event11th International Workshop on Machine Learning in Medical Imaging, MLMI 2020, held in conjunction with the 23rd International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2020 - Lima, Peru
Duration: 2020 Oct 42020 Oct 4

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12436 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference11th International Workshop on Machine Learning in Medical Imaging, MLMI 2020, held in conjunction with the 23rd International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2020
CountryPeru
CityLima
Period20/10/420/10/4

Keywords

  • Domain adaptation
  • Domain regularization layer
  • QSM
  • Region-based histogram matching
  • Segmentation
  • Substantia nigra

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint Dive into the research topics of 'Joint Appearance-Feature Domain Adaptation: Application to QSM Segmentation Transfer'. Together they form a unique fingerprint.

Cite this