Design and Analysis of Robust Image Registration Schemes using Machine Learning Algorithms

Abstract

Medical image registration plays a vital role in image-guided interventions to improve clinical decision-making, better visualization, and quantification of anatomical structures. Image registration is the process of transforming different sets of data into one coordinate system. Designing a universal framework for image registration is a challenging task due to the diversity of medical images and various forms of degradation that can occur during acquisition. A registration algorithm, rigid or deformable, aligns images within a common coordinate system to enhance their accuracy for diagnosis, monitoring, and treatment of medical conditions. This precise alignment facilitates clearer analysis and improved clinical decision-making. This research seeks to develop schemes that can effectively capture wide range of image registration scenarios by incorporating both rigid and deformable(non-rigid) registration techniques. Rigid registration is optimized using meta-heuristic methods to adjust rigid transformation parameters while deep learning is employed for feature extraction. In the deformable approach, a two-channel image from the moving and fixed images is processed through the U-Net model to generate a displacement field, while B-spline parameters predict and refine the warped image. The primary objective of this research is to develop a novel approach for rigid and non-rigid medical image registration applicable to both monomodal and multimodal modalities, utilizing machine learning and meta-heuristic optimization techniques. This study includes a thorough review of existing methods and algorithms for medical image registration, aiming to identify and address current gaps and limitations in the field. Firstly, area and feature-based methods are designed for monomodal and multimodal medical image modalities. The proposed approach employs teaching learning-based optimization (TLBO) to optimize rigid transformation parameters such as rotation and translation by considering mutual information (MI) maximization as an objective function. To address challenges in rigid transformation due to anatomical variations, a novel algorithm integrating both area-based technique such as TLBO and leveraging feature-based algorithm such as SURF are developed. This method combines the SURF framework for feature extraction with the RANSAC algorithm for feature refinement, followed by projective transformation for precise image registration. Afterwards, a robust registered image is generated by estimating the rigid transformation parameters based on inliers, thereby enhancing the registration accuracy. Secondly, a novel approach with hybridization of KPCA and TLBO for both monomodal and multimodal medical image registration is discussed. This approach starts with pre-processing such as noise removal and normalization to enhance image quality. Afterwards, contour extraction is carried out using Otsu's thresholding for multimodal images and a thresholding segmentation approach for monomodal images to create feature images and to determine ground truth translation parameters. Then, KPCA captures non-linear data mapping for rotation value determination, while TLBO optimizes the registration process by iteratively adjusting transformation parameters. Next, a deformable medical image registration leveraging an unsupervised learning algorithm and deep learning framework visual geometry group (VGG19) with TLBO and unified equilibrium optimizer (UEO) is proposed. The Vxnet (U-Net) model processes two-channel images to generate non-linear displacement vector field, while TLBO optimizes the registration process by iteratively adjusting transformation parameters. However, this deformable registration focuses on monomodal medical images. Afterwards, a novel image registration scheme for both monomodal and multimodal modalities leveraging UEO in contrast to TLBO optimization techniques in combination with VGG-19 is discussed. Then, features are extracted using VGG 19 from both reference and UEO-registered image. Dynamic inlier selection is applied to enhance the matching accuracy by effectively filtering correct matches (inliers) while minimizing outliers. Finally, thin plate spline interpolation is employed to determine the optimal transformation matrix for accurate alignment between reference and UEO-registered image. Lastly, an integration of the feature extraction technique SURF with VGG19 for anatomical and functional medical images is presented. The proposed scheme explores various feature extraction techniques namely SIFT, SURF, and ORB algorithms for detecting features. Then feature matching is performed using a brute force matcher, followed by outlier removal and projective transformation to generate the registered image. Despite the focus on feature extraction techniques, a novel image registration algorithm is introduced, combining deep learning with SURF for accurate image registration. The scheme enhances image quality through preprocessing and feature extraction using VGG19, followed by Euclidean distance calculations and dynamic inlier selection for refined alignment. Thin plate spline interpolation corrects misalignments, while the SURF algorithm and alpha-trimmed spatial correspondence minimize outliers. Finally, a homography matrix ensures robust alignment between sensed and VGG19-registered images. To validate the effectiveness of the proposed methodologies, experimental evaluations are conducted on publicly available and in-house clinical datasets for both monomodal and multimodal medical images. The results demonstrate the superiority and robustness of proposed techniques in terms of similarity metrics as compared to state of-the-art techniques.