Introduction
Since information from multiple medical imaging modalities is usually of complementary nature, proper extraction and registration of the embedded information and knowledge is essential to improve quality and safety of healthcare. Biomedical image registration enables the analysis and visualization of multimodality datasets simultaneously, and facilitates the integration and smart use of relevant anatomical and functional information. Because of its important role in clinical decision-making, operation planning, and image-guided surgery, brain image registration has been extensively studied, and numerous rigid algorithms and great progresses have been achieved. However, efficient elastic brain image registration is challenging and highly demanded for clinical applications, e.g., cranial image-guided surgery. We propose an innovative elastic and automatic registration method to improve the computational efficiency.
Methods
The approach consists of two steps. However, preprocessing step is required for multimodal brain image registration, for example, in the registration of MRI and PET images, morphological operations and Canny edge extraction algorithm need to be carried out to extract cerebral tissues from MRI images. Step 1 (efficient non-iterative affine registration): by using affine-invariant moment-based features, centroids, and major axes, affine parameters are directly derived by minimizing mean squared error (MSE) and time-consuming iterative optimization procedure is avoided. Step 2 (elastic registration based on active contour): active contours are energy-minimizing splines which can detect the closest contour of an object. The shape deformation of an active contour is driven by both internal energy and external energy. Firstly, contour is automatically extracted from study image as initial contour estimation of reference image. Because affine deformations have been corrected in step 1, this estimation can speed up the elastic registration convergence. and then, by iteratively applying external forces derived from reference image, the deformation field can be obtained, and elastic registration can be achieved.
Results
The approach has been validated by experiments on registering PET-MRI brain images and PET-PET images. By using IBM personal computer (Pentium 4, 3.0 GHz), the average computation time for 2-D images of 256*256 is about 68 seconds, mainly spent on elastic registration procedure. The initial contour estimation plays an important role in computational efficiency of elastic registration procedure. Different active contours influence registration performance and efficiency as well. Traditional expansion (ballooning) approach cannot result in high registration performance when initial contour estimation is much bigger than the real one. By using gradient vector flow (GVF) active contour, problem of concavity can be overcome and better registration performance can be achieved.
Conclusion
An efficient method is provided for elastic registration of biomedical brain images. Its clinical applications may include neurosurgery, minimally invasive procedure, disease monitoring, and treatment assessment.