人工智能核心550篇论文 医疗方向9.医学影像配准 1802.02604v3.pdfVIP

AnUnsupervisedLearningModelfor

DeformableMedicalImageRegistration

GuhaBalakrishnanAmyZhaoMertR.SabuncuJohnGuttag

MITMITCornellUniversityMIT

balakg@xamyzhao@msabuncu@guttag@

8AdrianV.Dalca

1

0MITandMGH

2adalca@

r

p

A

Abstractconstraintsontheregistrationmapping.Solvingthisop-

0timizationiscomputationallyintensive,andthereforeex-

2Wepresentafastlearning-basedalgorithmforde-tremelyslowinpractice.

]formable,pairwise3Dmedicalimageregistration.Cur-Incontrast,weproposeanovelregistrationmethodthat

Vrentregistrationmethodsoptimizeanobjectivefunctionin-learnsaparametrizedregistrationfunctionfromacollec-

Cdependentlyforeachpairofimages,whichcanbetime-tionofvolumes.Weimplementthefunctionusingacon-

s.consumingforlargedata.Wedefineregistrationasapara-volutionalneuralnetwork(CNN),thattakestwo-Dinput

cmetricfunction,andoptimizeitsparametersgivenasetofvolumesandoutputsamappingofallvoxelsofonevolume

[imagesfromacollectionofinterest.Givenanewpairoftoanothervolume.Theparametersofthenetwork,i.e.,the

3scans,wecanquickly