Research On Extraction Method Of Medical Image Lesion Area Based On Reinforcement Learning

Medical image processing plays an important role in medical diagnosis,clinical treatment,etc.However,the analysis and diagnosis of medical images are mainly completed by manual reading by radiologists.Constrained by many factors such as medical standards in backward areas,individual differences in patient pathology.There are great challenges in traditional manual reading,making it essential to implement an automated and accurate technique for extracting lesion areas in medical images.On the other hand,reinforcement learning is emerging in a wider range of fields,such as gaming,computer vision,etc.It has even surpassed human performance in some areas and is considered to be the way to implement artificial intelligence.Therefore,reinforcement learning has the potential to be applied to medical image processing.Based on the above reasons,this thesis explores the application of reinforcement learning in the task of extracting lesion areas from medical images.Specifically,this thesis builds models based on reinforcement learning for the brain tumor segmentation task of the BRATS dataset.Its main work is as follows:(1)Perform relevant analysis and pre-processing on the BRATS dataset.Pre-processing includes data standardization and data enhancement.In this regard,this thesis explores several mainstream image processing technologies,and verifies the effectiveness of Z-score standardization,random rotation,and random mirror flip through related experiments.(2)Through the exploration and analysis of Mask R-CNN series models,this thesis proposes an enhanced semantic segmentation framework,RSF,which is used to fuse semantic segmentation network and reinforcement learning network.The shared feature extraction backbone and semantic segmentation branch of this model can adopt any mainstream network structure,and its reinforcement learning branch can provide focus areas to improve the segmentation result.This thesis also conducted comparative experiments on the components of RSF,thus selecting the DenseNet and U-Net as the feature extraction backbone and upsampling structure of RSF,and confirmed that RSF can improve the segmentation accuracy of some categories.(3)On the basis of RSF,by analyzing the characteristics of the original 3D data slicing method,this thesis proposes a RL-based pseudo-3D attention network,RPAN.It introduces the prior knowledge to network in the form of attention through reinforcement learning,so as to improves the segmentation results for the current frame.This thesis also validates the structural rationality of the related modules of RPAN through a series of experiments,and the overall experimental results show that RPAN can achieve high accuracy while performing fast segmentation,which further confirms that the application of reinforcement learning in medical image processing has great potential and feasibility.