A novel 3D Registration Network with Intra-class alignment for cross-domain Neonatal Brain MRI segmentation

Bo Li, Xinge You, QINMU PENG, Jing Wang

Keywords: Neonatal Brain Segmentation, 3D Registration, Unsupervised Domain Adaptation, MRI

TL;DR: We propose RS-NET to make domain adaptation for the task of 3D brain MRI segmentation

Abstract: In neonatal brain Magnetic Resonance Image(MRI) segmentation, the model we train on dataset from specific medical institutions often fails to adapt clinical practice. The reason is that the clinical data(target domain) is always largely different from the training dataset(source domain) in terms of scale, shape, intensity distribution and so on. The registration network can transform the shape from the source domain to the target domain while the intensity transfer cannot be carried out. And current unsupervised domain adaptation(UDA) model which is based on the 2D GAN mainly focus on transferring the global intensity distribution to the target domain but cannot transform the shape and ignore the intra-class similarity. In this case, we propose a joint Registration and Segmentation Network(RS-NET) in which the two networks are jointly trained to perform both intra-class intensity transfer and shape transformation. An adaptive transfer layer(TL) is designed to intra-classly translate intensity from source to target which can make Registration focus on shape transformation and ignore the cross-domain intensity difference. Meanwhile, the segmentation network can adapt to the target domain through the transfer of registration network. The experiment is carried out on two databases which exist large differences in shape, size and intensity distribution between them. Our proposed method achieves state-of-the-art results in the compared UDA models for the 3D segmentation task. Source code (in Tensorflow) is available at: \href{url}{https://github.com/lb-whu/RS-NET/}.

Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.

Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.

Paper Type: both

Primary Subject Area: Segmentation

Secondary Subject Area: Transfer Learning and Domain Adaptation

Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.