Efficient and Robust Non-Rigid Least-Squares Rectification of Medical Images

One of the major problems facing medical imaging is the presence of geometric distortions inherent in an imaging technique. Image registration techniques are often used to correct for such geometric perturbations. Recently, it was proposed that the SSD cost function can be evaluated efficiently using the Fast Fourier Transform (FFT) to determine the optimal translation between two images. However, spatial distortions in medical images can be highly non-rigid in nature. This paper extends this efficient approach to allow for non-rigid alignment between two images through the use of patch correspondence and robust statistical model estimation techniques. This feature-based method is designed to be highly robust, making it suitable for aligning medical images with various forms of geometric distortions. Experimental results demonstrate high overall image rectification performance.