서지주요정보
Unsupervised medical image registration using cycle-consistent convolutional neural network = 주기적 일관성의 합성곱 신경망을 이용한 비지도 학습 기반의 의료영상 정합 연구
서명 / 저자 Unsupervised medical image registration using cycle-consistent convolutional neural network = 주기적 일관성의 합성곱 신경망을 이용한 비지도 학습 기반의 의료영상 정합 연구 / Boah Kim.
발행사항 [대전 : 한국과학기술원, 2019].
Online Access 원문보기 원문인쇄

소장정보

등록번호

8033617

소장위치/청구기호

학술문화관(문화관)B1층 보존서고

MBIS 19004

휴대폰 전송

도서상태

이용가능(대출불가)

사유안내

반납예정일

리뷰정보

초록정보

Image registration is one of the key processing steps for biomedical image analysis such as diagnosis of disease. Recently, deep learning based supervised and unsupervised image registration methods have been extensively studied due to its excellent performance in spite of ultra-fast computational time compared to the classical approaches. Most of the deep learning algorithms are designed to ensure diffeomorphism for the registration in calculating the deformation vector field. In this paper, based on the observation that a homeomorphic mapping between two topological spaces is as powerful as a diffeomorphism in ensuring the topology preservation and one-to-one mapping, we present a novel unsupervised medical image registration method that trains deep neural network to deform a 3D volume by homeomorphic mapping using a cycle-consistency. Using difficult multiphase liver registration tasks, we evaluate target registration error for the deformed images and demonstrate that the proposed method can provide accurate 3D image registration within a few seconds.

의료영상 정합은 영상 분석을 하기 위한 필수 단계 중의 하나이다. 기존의 정합 방법은 새로운 영상에 대해 정합 시간이 오래 걸리는 반면, 심층 학습 기반의 영상 정합 방법은 빠른 시간 안에 정합이 가능하여 최근 많은 연구가 진행되고 있다. 대부분의 심층 학습 기반 정합 방법은 변형 벡터장을 계산할 때 영상 정합에 대한 미분 동형 사상이 만족하도록 설계가 되어 있다. 이 논문에서는 두 개의 공간에 대한 위상 동형 사상 함수가 미분 동형사상처럼 일대일대응을 만족하고 형상을 보장한다는 것에 기반하여, 위상 동형 사상 함수에 의해 3차원 영상을 변형할 수 있도록 주기적 일관성을 이용하여 심층 신경망을 학습시키는 새로운 비지도 학습 기반의 영상 정합 기법을 제안하였다. 부 쓰는이는 다상의 간 영상을 정합하는 것에 제안한 방법을 적용하여 성능을 살펴보았고, 정량적 평가와 정성적 평가를 통해 제안한 영상 정합 기법이 빠른 시간 안에 정확한 3차원 영상 정합을 가능하게 함을 보였다.

서지기타정보

서지기타정보
청구기호 {MBIS 19004
형태사항 v, 38 p. : 삽화 ; 30 cm
언어 영어
일반주기 저자명의 한글표기 : 김보아
지도교수의 영문표기 : Jong Chul Ye
지도교수의 한글표기 : 예종철
학위논문 학위논문(석사) - 한국과학기술원 : 바이오및뇌공학과,
서지주기 References : p. 33-36
QR CODE

책소개

전체보기

목차

전체보기

이 주제의 인기대출도서

Example of2D slices taken from 3D liver CT volumes before and after injection of contrast agent. Images at different phases show various contrast and shape of liver and other organs.

The artificial neural network model. (a) The biological neural network that is composed of neurons. (b) The artificial neural network that imitates the biological neural network with nodes.

The graphs of activation functions. (a) A Sigmoid function. (b) A ReLU function (c) A leakyReLU function.

The example of convolution operation with stride 1 and padding 0 in convolutional layer. The value of orange region in output image is generated by convolution between the yellow box ofthe input and the kernel.

The example of pooling operation with stride 2 in pooling layer.

The network architecture of U-Net.

Buildingblocks for the neural network. (a) A building block for the CNN that takes input X, and generates output F(x), through several weighted layers. (b) A building block for ResNet. The block produces F(x)+x by shortcut that connects the input and output.

The model of conditional GAN. The generator takes an input that is concatenated with the latent variable 2 and specific information y. The discriminatoi takes the concatenation of real data 父 and the information y.

The model of cycleGAN. Both G and F denote generators that learn the distribution of input domains (X or Y) and map the distribution to the other domains (Y or X). Both Dx and Dy denote discriminators that distinguish between the real training data and the fake data resulting from the generators.

The overall medical image registration flows of the proposed method. Two convolutional neural networks (GA,GB) are used to take inputs by switching the order. Each CNN takes two vol- umes (A,B) and computes displacement vector field with three channels. Short and long dashed lines denote the moving images and fixed images, respectively. The 3D spatial transform function deforms the moving volume a

The diagram ofloss function structure in our proposed method. (a) The registration loss function, Cregist, computes difference in shape ofthe transformed image and fixed image. The cycle loss function, Ccycle, allows the displacement fields to return the deformed image to its original state as well as to transform the moving image. (b) The identity loss function, Cidentity, enable the network to b

The network architecture for generating a displacement vector field

Eight corners points of the cubic lattice with unit size around a point.

The architecture of polyphase U-Net.

The illustration of the process to extract slices only including liver in a volume

The operational parameters of rigid registration and non-rigid registration for the conven- tional method.

Results of multiphase liver CT registration. The diagonal images with red-box are original images, which are deformed to other phases as indicated by each row. Specifically, the (i.j),i=i element of the figure represents the deformed image to the i-th phase from the j-th phase original image.

Results of multiphase liver CT registration. The diagonal images with red-box are original images, which are deformed to other phases as indicated by each row. Specifically, the (i,j),i= j element of the figure represents the deformed image to the i-th phase from the j-th phase original image.

The 20 anatomic landmarks (yellow cross-shaped signs) ofthe liver in each abdominal CT images are marked by experts for the quantitative analysis.

The sample mean and sample standard deviation (std.) of target registration error (TRE) on the 50 test dataset. We compare our proposed method to the conventional method.

The graph for the mean values of TRE on the 50 test datasets. The cross-shaped points denote the sample mean ofthe TRE for each methods.

The mean values of major and minor axis for the tumor region in the 50 test dataset. The ground tumor sizes of the referenced portal phase images and deformed arterial/delayed phase images are produced by the experts. We compare our proposed method to the conventional method.

The average test time to deform one image into the fixed image. Our proposed method takes much less time to perform image registration compared to the conventional method.

Results ofthe ablation study. (a) Proposed method with only registration loss, (b) proposed method without identity loss, (c) proposed method without cycleloss, (d) proposed method. Each small images next to the CT images are the enlarged red boxes.

The example of failure cases for the proposed method