Advanced Feature Extraction and Outlier Detection for 3D Biological/Biomedical Image Registration

3D image registration is essential in computer vision, medical imaging, and robotics. By aligning images from different perspectives into a single coordinate system, this approach provides a consistent viewpoint for analysis. Using accurate image alignment, we may compare, evaluate, and integrate data from different contexts. This paper describes a new method to register 3D or z-stack microscopy and medical image. It uses a hybrid of traditional and deep learning methods for feature extraction and adaptive likelihood-based methods for finding outliers. The proposed method uses the Scale-invariant Feature Transform (SIFT) and the Residual Network with 50 layers (ResNet50) to extract effective features to obtain precise and accurate representations of image contents. The registration approach also relies on the adaptive Maximum Likelihood Estimation SAmple Consensus (MLESAC) method, which optimizes outlier detection and increases noise and distortion resistance to improve the efficacy of these combined extracted features. This concatenation approach demonstrates robustness, flexibility, and adaptability across a variety of imaging modalities, enabling the registration of complex images with higher precision. Results show that the proposed algorithm outperforms commonly used registration methods, including SIFT, KAZE, Oriented FAST and Rotated BRIEF (ORB), and also registration software tools such as bUnwarpJ, and TurboReg. The algorithm's effectiveness is evaluated in terms of Mutual Information (MI), Phase Congruency-Based (PCB), and Gradient-Based Metrics (GBM). These metrics are applied to two types of datasets, including a brain scan dataset and 3D serial sections of multiplex microscopy image datasets.