Lesion Segmentation Based On Multi-scale Of Reinforcement Learning And Multi-task With Research On Adaptive Multi-organ Few-shot Segmentation

Medical image is widely used in various fields such as clinical auxiliary diagnosis,intraoperative guidance,and early disease screening.It is an important reference method for doctors in diagnosis and greatly improves the accuracy of disease diagnosis.Recently,with the rapid development of deep learning,artificial intelligence has been widely used in the field of image processing,especially in the field of medical imaging.A large number of segmentation and classification detection algorithms with excellent performance have been proposed and shown excellent performance on a large number of public datasets.The application of these algorithms greatly reduces the doctor’s repetitive work,relieves the doctor’s burden of reading medical images,and improves the doctor’s work efficiency and the patient’s treatment effect.Based on the actual problems of clinicians,this paper proposes corresponding deep learning segmentation methods for different problems in lesion area segmentation and multi-organ segmentation to improve the segmentation result in different segmentation tasks and assist doctors in clinical diagnosis.The methods are as follows:1.A method for segmentation of ankylosing spondylitis lesions based on reinforcement learning multi-scale neural network is proposed to solve the problems in the segmentation of lesions due to irregularities in the scale,size,and location of the lesions,which are caused by blurred and diffused lesions.The designed network can improve the segmentation performance of multi-scale lesion.This method designs a multi-scale convolution module HMS composed of different dilated ratios and different sizes of convolution kernels,and uses this module to construct a multi-scale segmentation network that is used to deal with large differences in the shape and size of the lesion area.Meanwhile,a hard sample mining method based on the data augmentation sub-network of reinforcement learning is designed,the constructed multi-scale segmentation network and the data augmentation sub-network are connected in parallel to build the segmentation model LHR-Net,which completes the segmentation of the multi-scale and diffuse fuzzy lesion area.The experiment results prove that the network has good segmentation performance for multi-scale lesions.2.A method for segmentation of lesions based on category priors is proposed to solve the problem of large differences in the segmentation results of lesions of different severity levels.In this research,lesion areas of different severity levels behave differently,and the network conducts unified training for different types of lesions,resulting in differences in segmentation effects.In order to improve the overall performance of the segmentation network,this method introduces the lesion category information of the data into the segmentation network,trains the multi-task network to predict the segmentation result and category of the lesion area of the data simultaneously,by introducing category supervision information to assist the network to extract more robust feature information for different types of lesions to improve the overall performance of lesion area segmentation,the experiment results show that the proposed method improves the results of different types of lesion segmentation.3.A self-supervised few-shot adaptive multi-organ segmentation method is proposed to solve the query set segmentation result,due to insufficient support set supervision information and lack of query set prior information in the multi-organ segmentation task when using the few-shot segmentation method,the network first uses the encoder to extract the feature information of the target area of the support set and the query set,then generates the probability map of the query set segmentation based on the maximum global similarity,which is used to describe the global similarity between the support set and the query set Using the generated probability map as the prior probability map of the query set to further guide the segmentation task of the query set.By calculating the global similarity probability map as the segmentation prior information of the query set,this method can make full use of the guidance information provided by the support set to improve the segmentation effect of the query set.Experiments have proved that this method can effectively improve the segmentation effect of multiple organs.