Registration of 3D serial MR brain images


London University (March, 2001)

Mark Holden,

Radiological Sciences, GKT, Guy's Hospital, London, SE1 9RT, UK.

e-mail: mark.holden@kcl.ac.uk

Abstract

This thesis investigates automatic registration to detect and quantify small structural changes in brain anatomy; and also to rectify scanner geometrical distortion. Five treated growth hormone (GH) patients, six normals and a cube phantom were imaged serially at 0, 3 and 6 months using a Siemens 3D FLASH T1 weighted gradient echo MR sequence. Eight voxel similarity measures were quantitatively evaluated for rigid-body registration of the serial brain scans by measuring registration consistency. The serial images were rigidly registered, with and without brain pre-segmentation and scaling error correction (obtained from the nine degrees of freedom registration of phantom images), and interpolated using a sinc kernel to produce sixty-six difference images. These difference images were randomised and visually assessed for structural change by two neuro-radiologists. Structural change was also quantified with a non-rigid registration algorithm based on a tensor product of cubic B-splines and optimisation of normalised mutual information. The geometrical distortion of a Philips ACS2 scanner was investigated with a phantom consisting of a 3D matrix of 427 spherical reference structures. Distortion was determined by point-based non-rigid registration, with a fourth order polynomial, of imaged and modelled sphere centre point-sets. The evaluation of similarity measures indicated that mutual information was significantly (p < 0.01, n=33) more consistent than the other six measures. The blinded visual assessment of difference images indicated a significant (p < 0.01) increase in brain/CSF volume ratio for five GH patients compared to six normals. Quantitative measurements gave a mean ventricular volume change over 6 months of: -1.2cc for GH patients compared to +0.18cc for normals with a significant (p=0.013) difference between the two groups. There was high correlation (rho ~ 0.7, n=11) of the quantitative and visual measurements. Distortion measurements indicated large shear and second order components and negligible distortion change over 11 weekly scans.

Thesis in PDF format:

  • Thesis in pdf (available to researchers on request)
    Mark Holden research.mholden.org
    Last modified: 4 May, 2001.