Medical image registration based on feature and mutual information

Abstract

Image registration based on mutual information (MI) has been widely used in remote sensing data analysis, computer vision, medical image disposal and other fields. But the mutual information is calculated by the joint histogram of two images which don't take into account the space-position relationship of the image pixel, so the registration precision will be degraded. Aiming at the lack of mutual information registration, a new method of medical image registration based on mutual information of multi-scale Harris corner and feature points was given. It effectively improves sensitivity of noise and the situation of being trapped into a local minimum to MI registration.

REAL-X—Robot Open-Ended Autonomous Learning Architecture: Building Truly End-to-End Sensorimotor Autonomous Learning Systems

“…The capacity to have long interactions with objects would probably require close-loop policies and haptic sensors, solutions not explored here. Since hardwiring behaviours is not allowed in REAL-X (e.g., as done in the system presented in [64]), the attainment of effects similar to those of the grasping reflex might be possibly achieved with novelty or surprise intrinsic motivations used to mark the saliency of the hand-object interactions [65]: as an example, [46] used mutual information between agent state (proprioception) and environment state (object positions) to achieve good results in a pick and place scenario.…”

“…These systems can potentially be applied to the REAL benchmark but this is prevented by the limitations discussed below. The system proposed in [46] is also relevant as it trains a robotic agent in a pick and place scenario ('OpenAI FetchPickAndPlace') without using an external reward, but using an intrinsic reward based on the mutual information between the agent state and the environment state. However, in the used scenario the robot does not use a camera but has access to the position of objects.…”

Maximum Entropy Population-Based Training for Zero-Shot Human-AI Coordination

“…From a Bayesian perspective, when we do not have a prior on what the human policies look like, we should train the AI agent to be robust and capable of collaborating with a diverse set of policies (Murphy 2012). One popular approach towards robust AI agents is through maximum entropy reinforcement learning (Ziebart et al 2008;Ziebart 2010;Fox, Pakman, and Tishby 2015;Haarnoja et al 2017Haarnoja et al , 2018b, and many previous works leverage it as a means of encouraging exploration (Schulman, Chen, and Abbeel 2017;Haarnoja et al 2018b) or skill discovering (Eysenbach et al 2019;Zhao et al 2021). However, obtaining a diversified population through entropy maximization is still subjective to research.…”

Two-dimensional forward-looking sonar image registration by maximization of peripheral mutual information

“…Pluim et al 23 included spatial information by combining mutual information with a term based on the image gradient. On the other hand, Zhao et al 28 adopted the Rényi entropy in place of Shannon entropy to reduce the effects of local extremes to the registration function. Voronov and Tashlinskii 29 compared the gradient of different entropies, where the probability function is described by the Gauss-Parzen window function.…”

“…The joint and marginal probabilities that are used in calculating the mutual information take into account only the relationships between corresponding individual pixels. [22][23][24]27,28 However, when modelling the sonar image with a Markov random field, 12,30,31 it is found that pixel intensities depend on the neighbouring pixels. Such a positive correlation may be introduced by the reverberation of the seabed and the scattering effects of the water on the acoustic waves, which demonstrates that it is necessary to include the neighbourhood information in registering the 2D sonar image pairs.…”