July 15, 2011 Registering retinal images Babak Ghafaryasl Universitat Pompeu Fabra Csaba Molnár University of Szeged Antonio R. Porras Universitat Pompeu Fabra Arie Shaus Tel Aviv University http://www.inf.u-szeged.hu/projectdirs/ssip2011/teamG Overview Vessel enhancement Vascular tree extraction Feature extraction Registration Vessel enhancement “Multiscale vessel enhancement filtering”, Frangi et al, 1998 - Scale Space representation - Local image descriptors - Eigenvalues of Hessain (2nd derivative) matrix Tubular, plate-like and spherical structures Vascular tree extraction Original images Vessel enhancement Thresholding + Skeletonization + Largest connected components 7 6 8 L2 9 5 1 L1 4 2 3 10 11 12 L3 From bifurcation point to bifurcation structure… “Feature-Based Retinal Image Registration Using Bifurcation Structures”, Chen & Zhang, 2009 x [ L1, L2 , L3 , 1, 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 , 11, 12 ] But… • L’s are normalized to sum up to 1. • The α triplets sum up to 360. Therefore we can remove some redundancy. x [ L1 , L2 , 1 , 2 , 3 , 5 , 6 , 7 , 10 , 11 ] We can measure a distance between such structures! From bifurcation structures to registration… Step 1: Find bifurcation structures in both images. Step 2: Find the best match between two bifurcation structures. The match between 4 points (3 are enough) determines the affine transformation. Step 3: Find next best matches (taking the transformation into account); refine the affine transformation with more points. Results: feature extration All candidates Matching candidates Vessel registration Results: retinal registration (I) Results: retinal registration (II) Results: bad news... Questions?