Spectral clustering to model deformations for fast multimodal prostate registration

This paper proposes a method to learn deformation parameters off-line for fast multimodal registration of ultrasound and magnetic resonance prostate images during ultrasound guided needle biopsy. The registration method involves spectral clustering of the deformation parameters obtained from a spline-based nonlinear diffeomorphism between training magnetic resonance and ultrasound prostate images. The deformation models built from the principal eigen-modes of the clusters are then applied on a test magnetic resonance image to register with the test ultrasound prostate image. The deformation model with the least registration error is finally chosen as the optimal model for deformable registration. The rationale behind modeling deformations is to achieve fast multimodal registration of prostate images while maintaining registration accuracies which is otherwise computationally expensive. The method is validated for 25 patients each with a pair of corresponding magnetic resonance and ultrasound images in a leave-one-out validation framework. The average registration accuracies i.e. Dice similarity coefficient of 0.927 ± 0.025, 95% Hausdorff distance of 5.14 ± 3.67 mm and target registration error of 2.44 ± 1.17 mm are obtained by our method with a speed-up in computation time by 98% when compared to Mitra et al. [7].