Cluster Based Medical Image Registration Using Optimized Neural Network

Introduction

Image registration or alignment and matching of two or more images, establishes a one-to-one spatial correspondence of a single 2-D/3-D scene or several similar scenes captured at different time instants or from various viewpoints or by different sensors. In image processing this is one of the important steps used in a variety of applications including remote sensing and cartography, autonomous navigation, robot vision, and medical imaging to mention a few. It is a powerful tool for integrating or fusing image data collected from different sensors (mono-modal or multi-modality), tracking the temporal evolution (changes in images taken at different times), making inter-patient comparisons, reconstructing 3-D (volumetric) images from multiple 2-D (planar) images, etc. When one image is registered to another image, the latter is typically referred to as a reference image, and the former is called a target or sensed image. Medical images are very important for diagnosis, treatment, and supervising disease progression. The term ‘medical image’ spreads over a vast area of different types of images. Now-a-days medical researchers used medical images to investigate disease processes and to understand different developments like a tumor. Multiple images are acquired of common subjects at different times, or from different imaging modalities. Comparing two or more unregistered images can lead to incorrect diagnostic conclusions. Medical image registration generally deals with the technique to align two or more images of different modalities like inter-modality or intra-modality. Different topographic medical images like computed tomography (CT), magnetic resonance imaging (MRI), single photon emission computed tomography (SPECT), and positron emission tomography (PET) are generally used for research in image registration. Computer-aided diagnosis (CAD) systems use image registration to investigate how human anatomy is altered by disease, age, gender, handedness, and other clinical or genetic factors. Over the last decade, researches on automatic rigid registration methods of medical images have been widely developed. Rigid registration generally deals with some popular global geometric transformations include similarity, affine transformation, perspective projection, and polynomial models etc. Affine geometric transformation was used by Scale-invariant feature transform (SIFT) introduced by Lowe (Lowe, 2004). Multiple point-wise correspondences between local areas in two images determine the transformation parameters. Different retinal images collected with the 1-day time difference is analyzed using this method to estimation transformation parameters. Affine transformation deals with translation, rotation, scaling, and skewness of a target with respect to a reference. For negligible deformation in the image the affine transformations are sufficient. To address the deformable nature of a medical image like different organs, tissues etc. nonrigid registration methods have been developed. However, the global mapping is unable to capture intrinsically local large deformations of anatomical structures. Thus, frequently more flexible elastic transformations that locally warp a target to align with a reference image are needed. Most popular such transformations include large deformation models (diffeomorphisms), radial basis functions (RBF), physical continuum models (viscous fluids). Soft Computing techniques like GA (Genetic Algorithm), ANN (Artificial Neural network) etc. are also used in medical image registration. A powerful advantage of soft computing is the complementary nature rather than competitive. For this reason more than one soft computing technique is used collectively named hybrid technique. Hybrid soft computing models have been applied to a large number of classification and prediction. In this chapter a hybrid technique is described to register images of same modality. Here weight optimization of neural network is done by GA.