A Point Cloud Reconstruction Model Of Brains Based On Generative Adversarial Strategies And Graph Convolutional Networks

In recent years,as surgical approaches continue to evolve,cutting-edge medical surgical techniques such as minimally invasive surgery and robotically guided surgical interventions have been increasingly used in brain surgery.New generations of biotechnology,such as brain-computer interface(BCI),are possible thanks to similar surgical techniques.However,this type of surgery also places new demands on the ability to collect information in real time during the operation.As surgeons are unable to directly observe the surgical targets and objectives during the procedure,their clinical experience is often not effectively utilised.As a result,a variety of intraoperative imaging navigation techniques have been proposed to alleviate this problem,as imaging is comprehensive and clear,and Magnetic Resonance Imaging(MRI)navigation is currently the main intraoperative navigation technique.However,there are many problems with today's MRI navigation techniques: 1)MRI navigation is not time efficient.The number of MRI pixels for complete imaging is in the order of a million,and the computation and processing process is greatly deprived of real time.2)Since most MRI systems in the operating room are mid-field MRI systems,the signal-to-noise ratio of the images obtained from functional imaging will not be very high and the position determined will not be very accurate.3)MRI cannot be visualised directly,and the complex step of extracting a 3D model of the surgical organ from MRI reduces the degree of automation of the navigation automation of software and hardware.3D reconstruction methods of organs combined with deep learning techniques are proposed in order to find some accurate and controllable visual enhancement methods to improve MRI navigation techniques and obtain more complementary information.However,no work has yet explored this technique in the field of brain surgery,and the aim of this paper is to propose a 3D reconstruction method for the brain with accuracy and robustness to fill this gap.In this paper,a tree-graph convolutional generative adversarial network model for minimally invasive brain surgery scenarios is designed with powerful feature adaptive extraction and learning capabilities to reconstruct accurate 3D point clouds of the brain at short time delays.It applies a tree-structured graph convolutional network in the generator part to fully exploit the graph node correlation of the point cloud and to improve the accuracy of the generation.The adversarial generation strategy ensures that Nash equilibrium can be effectively achieved by multiple modules and that the model can more effectively approximate the generated distribution.In addition,existing conventional medical images such as Computed Tomography(CT)and Magnetic Resonance Imaging(MRI)often suffer from at least one of two problems during surgery: firstly,they often suffer from a variety of possible visual contaminants due to lighting constraints and outside the surgical plan(e.g.local hemorrhage),they are often incomplete.Secondly,their resolution is not always adequate for the exponentially increasing surgical refinement required.To address the first problem,this paper designs a hierarchical shape perception network combining a selfattentive mechanism and a generative adversarial strategy to reconstruct the 3D point cloud structure of the brain from incomplete 2D images,and various modules are designed to ensure as much confidence as possible in the missing completions.For the second problem,a pyramidal shape-aware network with two-stage upsampling generation capability is designed to reconstruct the regular low-density point cloud and then proceed to generate a high-density point cloud with a larger number of points and finer details.A hybrid loss function combining adversarial loss,Kullback-Leibler scatter constraint,reconstruction loss,and perceptual loss is also purposefully designed for each task to improve the training efficiency of the model.Finally,the paper also demonstrates the good performance of the designed network models in various studies through extensive stereotypical,quantitative,ablation,and comparison experiments.