Joint segmentation and deformable registration with the presence of the noise using active contour and linear curvature

Medical image processing requires image segmentation and image registration as important diagnostic tools. Image segmentation aims to split structures of interest into distinct components. Meanwhile, image registration finds an ideal transformation between given images. Despite their interrelationship, both tasks are often performed separately. As a result, it faces several problems when undertaken independently. Therefore, the joint models between segmentation and registration have advantages over the separate tasks, particularly when dealing with noisy images. This paper reviewed and compared three existing variational joint segmentation and registration (JSR) models based on active contour without edges and linear curvature. The models are tested on 2D mono-modal and multi-modal images with the presence of white Gaussian noise. The Dice coefficient metric and the registration performance measure are used to evaluate the performance of the models. Numerical results showed that the JSR model using the modified sum of the squared difference outperformed the other two JSR models for noisy mono-modal images. However, the JSR model using the modified normalized gradient fields performed better than the JSR model based on the modified mutual information for noisy multi-modal images.