人工智能核心550篇论文 医疗方向9.医学影像配准 1902.05020v3.pdfVIP

1

Unsupervised3DEnd-to-EndMedicalImage

RegistrationwithVolumeTweeningNetwork

ShengyuZhao,TingfungLau,JiLuo,EricI-ChaoChang,YanXu

Abstract—3Dmedicalimageregistrationisofgreatclinicalim-

portance.However,supervisedlearningmethodsrequirealarge

amountofaccuratelyannotatedcorrespondingcontrolpoints

(ormorphing),whichareverydifficulttoobtain.Unsupervised

9learningmethodseasetheburdenofmanualannotationby

1exploitingunlabeleddatawithoutsupervision.Inthispaper,

0weproposeanewunsupervisedlearningmethodusingconvolu-(a)fixed(b)moving(c)deformation(d)warped

2tionalneuralnetworksunderanend-to-endframework,VolumeFigure1:Illustrationof3Dmedicalimageregistration.Givenafixedimage(a)anda

tTweeningNetwork(VTN),for3Dmedicalimageregistration.movingimage(b),adeformationfield(c)indicatesthedisplacementofeachvoxelinthe

cWeproposethreeinnovativetechnicalcomponents:(1)Anfixedimagetothemovingimage(representedbyagridskewedaccordingtothefield).

Animage(d)similartothefixedonecanbeproducedbywarpingthemovingimage

Oend-to-endcascadingschemethatresolveslargedisplacement