A Novel Approach for Pathology Detection using CNN based Image Registration Techniques

Abstract

Automatic detection of pathology in images can help in reducing the workload of pathologists and speed up the diagnosis. Advancements in medical imaging technologies have enabled high-quality visualization of tissue structures for anatomical and pathological examinations. This paper is aimed at developing fast and reliable systems for pathology localization towards improving diagnostic accuracy. To develop a robust and accurate image registration algorithm for pathology localization in MRI brain images. A two-stage unsupervised end-end deep learning model called Rigid Transform and B-spline based Convolutional Neural Network (RBCNN ) is proposed for registration of MRI images for brain tumor localization. This paper exploits the rigid transform in the first stage, to extract the rigid transform parameters for pre-registration, resulting in a coarse alignment of the images. Then the B-spline transform is used in the second stage to align the image further to obtain the optimal alignment. This paper exhibits promising results with low computation time and robustness compared to state-of-the-art methods. RBCNN can be integrated with existing image segmentation algorithms to segment the tumors. Further, it can be used as a backbone in tumor detection networks.