Imperial College London


Faculty of EngineeringDepartment of Computing

Head of Department of Computing



+44 (0)20 7594 8333d.rueckert Website




568Huxley BuildingSouth Kensington Campus






BibTex format

author = {Li, Y and Alansary, A and Cerrolaza, J and Khanal, B and Sinclair, M and Matthew, J and Gupta, C and Knight, C and Kainz, B and Rueckert, D},
publisher = {Springer Verlag},
title = {Fast multiple landmark localisation using a patch-based iterative network},
url = {},

RIS format (EndNote, RefMan)

AB - We propose a new Patch-based Iterative Network (PIN) for fast and accuratelandmark localisation in 3D medical volumes. PIN utilises a ConvolutionalNeural Network (CNN) to learn the spatial relationship between an image patchand anatomical landmark positions. During inference, patches are repeatedlypassed to the CNN until the estimated landmark position converges to the truelandmark location. PIN is computationally efficient since the inference stageonly selectively samples a small number of patches in an iterative fashionrather than a dense sampling at every location in the volume. Our approachadopts a multi-task learning framework that combines regression andclassification to improve localisation accuracy. We extend PIN to localisemultiple landmarks by using principal component analysis, which models theglobal anatomical relationships between landmarks. We have evaluated PIN using72 3D ultrasound images from fetal screening examinations. PIN achievesquantitatively an average landmark localisation error of 5.59mm and a runtimeof 0.44s to predict 10 landmarks per volume. Qualitatively, anatomical 2Dstandard scan planes derived from the predicted landmark locations are visuallysimilar to the clinical ground truth.
AU - Li,Y
AU - Alansary,A
AU - Cerrolaza,J
AU - Khanal,B
AU - Sinclair,M
AU - Matthew,J
AU - Gupta,C
AU - Knight,C
AU - Kainz,B
AU - Rueckert,D
PB - Springer Verlag
SN - 0302-9743
TI - Fast multiple landmark localisation using a patch-based iterative network
UR -
ER -