ComputerizedMedical Imagingand Graphics36 (2012) 130– 138 Contentslists availableat SciVerseScienceDirect Computerized Medical Imaging and Graphics j o u r n a l h o me p a g e : w w w . e l s e v i e r . c o m / l o c a t e / c o m p m e d i m a g Constrainedreversediffusion for thick slice interpolation of 3D volumetric MRI images Aleˇ s Neubert a,∗ , Olivier Salvado a,1 , Oscar Acosta a,b,c,2 , Pierrick Bourgeat a,3 , Jurgen Fripp a,4 a CSIRO ICT Centre, TheAustraliane-HealthResearch Centre, Brisbane, Australia b Université deRennes 1, LTSI, Rennes, F-35000, France c INSERM, U1099, Rennes, F-35000, France a r t i c l e i n f o Articlehistory: Received6 April 2011 Receivedin revisedform 8 August 2011 Accepted10 August 2011 Keywords: Imageinterpolation Magnetic resonanceimaging Three-dimensional(3D) imaging Reversediffusion B-Splines Resolutionenhancement a b s t r a c t Due to physical limitations inherent in magneticresonanceimaging scanners, three dimensional volu- metric scansare often acquired with anisotropic voxel resolution. We investigateseveral interpolation approachesto reduce the anisotropy and present a novel approach– constrainedreversediffusion for thick slice interpolation. This technique was comparedto common methods: linear and cubic B-Spline interpolation and a technique basedon non-rigid registrationof neighboring slices. The methods were evaluated on artificial MR phantomsand real MR scansof human brain. The constrainedreversediffusion approachdeliveredpromising results and providesan alternativefor thick slice interpolation,especially for higher anisotropyfactors. © 2011 Elsevier Ltd. All rights reserved. 1. Introduction The acquisition of magnetic resonance(MR) images inherently involves a trade-off between image quality and scan time. Tra- ditionally MR images are acquired as high resolution 2D slices with relatively large slice thicknesses in comparison to the in-slice resolution, and sometimes with inter-slice gaps. For images with thick slices,blurring occurswhich challenges most post-processing techniques. Imageswith inter-slice gapsmiss information and can increase the number of false negativesin segmentationresulting in under-estimation of pathologies.Volumetric imaging or three- dimensional (3D) MR sequences have been used to reduce this problem and provide images where each sample is the result of a full volumetric voxel excitation rather than excitation of one selectedpoint. Therefore, the collected signal carries information from the full imagedvolume instead of severalchosenplanes.Par- tial voluming effects(PVE) occur when the single voxel is a mixture ∗ Corresponding author at: The Australiane-Health Research Centre, CSIRO, Level 5 UQ Health Sciences Building 901/16,Royal Brisbaneand Women’s Hospital, Her- ston, Queensland4029,Australia. Tel.: +617 3253 3638; fax: +617 32533690. E-mail addresses: ales.neubert@csiro.au (A. Neubert), olivier.salvado@csiro.au (O. Salvado),oscar.acosta@univ-rennes1.fr (O. Acosta), pierrick.bourgeat@csiro.au (P. Bourgeat), jurgen.fripp@csiro.au (J. Fripp). 1 Tel.: +617 3253 3658. 2 Tel.: +33223235334. 3 Tel.: +617 3253 3659. 4 Tel.: +617 3253 3660. of signals from several tissue types, blurring the exact boundaries betweenneighboringtissuesand consequently reducethe segmen- tation precision [1]. Volumetric acquisitions will increasethe PVE if the resolution is not adaptedto the structures being imaged.By cutting the original voxel into smaller ones in interpolation, one desirable property would be to apportion the intensity among the smaller volumes while conserving its integral which reflects the composition of the underlying tissues. Many visualization, image processing and segmentationalgo- rithms rely on having high resolution isotropic voxels that exhibit minimal PVE. To achievethis on 2D imageswith thick slices, addi- tional slices are calculated(interpolated) using the set of acquired 2D slices and various image post-processingproceduresto gener- ate a high resolution image [2,3]. As the nature of the volumetric MR scans is substantially different from MR images with thick slices, the performance of classical interpolation techniques may vary depending on the type and resolution of the images we are processing. There have been numerous interpolation techniques with var- ious characteristics and effectiveness presented for MR images [2–4]. Classical mathematical techniques use a set of basis func- tions to estimate values in-between the slices. Nearest neighbor, linear and cubic B-Spline functions are the standard types of such interpolation approaches that are commonly used in med- ical imaging [3]. There are now several alternative approaches adapted specifically for medical images which attempt to con- sider the anatomical objects being interpolated and take into account their shape and appearance. Most notably, shape-based 0895-6111/$– seefront matter ©2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.compmedimag.2011.08.004