INTEGRATING STRUCTURAL AND DIFFUSION MR INFORMATION FOR OPTIC RADIATION LOCALISATION IN FOCAL EPILEPSY PATIENTS Pankaj Daga ⋆ , Gavin Winston † , Marc Modat ⋆ , M. Jorge Cardoso ⋆ , Jason Stretton † , Mark Symms † , Andrew W. McEvoy ⊕ , David Hawkes ⋆ , John Duncan † , Sebastien Ourselin ⋆ ⋆ Centre for Medical Image Computing, University College London, London, UK † Department of Clinical and Experimental Epilepsy, UCL Institute of Neurology, London, UK ⊕ National Hospital for Neurology and Neurosurgery, London, UK ABSTRACT Current state of the art neurosurgical systems do not ex- ploit the complementary information provided by struc- tural and diffusion MRI when aligning pre-operative and intra-operative images. We propose a multivari- ate registration scheme where structural and fractional anisotropy data are combined in a single similarity mea- sure. We formulate the normalised mutual information expression for the multichannel scheme and compute its analytical derivative. The method was validated using both a numerical phantom and clinical data using pre and post-operative images from patients who had under- gone surgery for treatment of refractory focal epilepsy and shows correlation between visual field deficit and predicted damage to the optic radiation. This work could be of significant utility in image-guided interven- tions and facilitate effective surgical treatments. Index Terms — multivariate registration, optic ra- diation localisation, interventional MR 1. INTRODUCTION Around one-third of patients with focal epilepsy are re- fractory to treatment with anti-epileptic drugs. Ante- rior temporal lobe resection is an effective treatment for such patients with refractory temporal lobe epilepsy [1]. However, a careful balance has to be established between obtaining seizure control and minimising the chance of We are grateful to the Big Lottery Fund, Wolfson Trust and National Society for Epilepsy for supporting the NSE MRI scanner. This work was undertaken at UCLH/UCL who received a propor- tion of funding from the Department of Healths NIHR Biomedical Research Centres funding scheme. Acquisition of patient scans was funded by a Wellcome Trust Programme Grant (WT083148), Pankaj Daga was funded by joint Cancer Research UK/EPSRC grant(C1519/A10331) and Gavin Winston was funded through a Clinical Research Training Fellowship from the Medical Research Council (MRC grant number G0802012). causing new morbidity. Significant morbidity can result from damage to eloquent grey matter regions of the brain or critical white matter tracts. One major source of mor- bidity in these cases is the damage to the optic radiation during the intervention. Hence, accurate localisation of optic radiation is critical in improving the surgical out- come for patients undergoing anterior temporal lobe re- section. Current neurosurgical protocols require acquisition of images using multiple modalities, for example, T1- weighted, T2-weighted and diffusion tensor (DT) mag- netic resonance (MR) images. These images can often provide unique and complementary information about the underlying tissue. Structural MR images can cap- ture information at the interfaces between the different brain tissues while DT-MR images can provide informa- tion about organizational structure of the white matter fibre bundles. The fusion and presentation of informa- tion to the surgeons from these images could facilitate critical surgical decisions and improve the ultimate sur- gical outcome. The current commercial image-guided neurosurgical systems do not perform non-rigid mapping between the pre and intra-operative images due to the inherent time constraints in a neurosurgical procedure. They also do not exploit the shared information between these complementary images. This limitation reduces their accuracy in providing accurate localisation of ob- jects of interest. Archip et al [2] demonstrated a neuro- surgery system that performs non-rigid registration and can visualise data from various imaging modalities. How- ever, they do not exploit the shared information between these images in their registration scheme. Park et al [3] used a multiple-channel demons algo- rithm to normalise DT-MRI data and applied it to cre- ation of group diffusion tensor atlas. They performed registration using different combination of channels to si- multaneously map structural and diffusion images. Sim- ilarly, Avants et al [4] presented a multivariate approach