A 3D/2D Medical Image Registration Method Based on Deep Reinforcement Learning

Abstract

The invention discloses a 3D / 2D medical image registration method based on deep reinforcement learning, which first acquires 2D and 3D medical images and performs preprocessing to obtain multi-resolution reference images and floating images; then builds a deep reinforcement learning network, Project the 3D images to be registered before operation in a specific direction to obtain DRR images with different transformation parameters, then calculate the similarity measure between each DRR image and the 2D reference image, and finally select the DRR image with the highest similarity measure value, and finally use A deep reinforcement learning network performs image registration on the selected DRR images.

Description

technical field [0001] The invention belongs to the technical field of image registration, and more specifically relates to a 3D / 2D medical image registration method based on deep reinforcement learning. Background technique [0002] Image-guided surgery in clinical operations usually requires obtaining 3D images of human body lesions before the operation to help doctors understand the patient's condition and formulate surgical planning. The key to surgery is to accurately establish the spatial position relationship between the preoperative 3D image and the intraoperative 2D image, that is, to register the preoperative 3D image and the intraoperative 2D image. [0003] Current registration methods can be divided into grayscale-based methods, feature-based methods, and deep learning-based methods. The feature-based registration method can be divided into external feature-based and internal feature-based. Based on external features, medical imaging is performed after implanti...

AI Technical Summary

Problems solved by technology

In general, feature-based registration ignores a large amount of image information, resulting in low registration accuracy and requiring human intervention

The grayscale-based registration algorithm uses the similarity measure to quantify the similarity between the 2D image generated by projection and the intraoperative 2D image, and uses the optimization algorithm to iteratively search for the optimal similarity measure value to represent the best registration state of the image. , which is the optimal spatial transformation parameter, the use of more grayscale information means more processing data, complex calculation, longer registration time, and poor real-time performance

Shun Miao proposed to use a deep regression network to directly predict 2D / 3D registration transformation parameters, but the preprocessing steps are complex, the network structure is lengthy, a large amount of data is required, and the end-to-end direct prediction of transformation parameters is not conducive to ensuring registration accuracy

Images