Design On Automatic Target Delineation Algorithm For Tumor Radiotherapy

Radiotherapy is one of the main methods in clinical cancer treatment,which usually includes image collection,target delineation,radiotherapy planning and implementation.Among them,the delineation of tumor target is an important prerequisite for the formulation and implementation of radiotherapy plan,which is usually completed manually by physicians.However,the manual delineation process is time-consuming and laborious,and there are subj ectivity and limitations of human eyes,which affect the efficiency and accuracy of tumor target delineation,and then affect the formulation and implementation of radiotherapy plan.The delineation of tumor target can be regarded as a binary pixel classification problem,that is,each pixel in the image can be divided into two categories according to whether it is a tumor target volume or not.With the continuous development of deep learning technology in the field of digital image processing and the updating of computing equipment,tumor intelligent analysis and diagnosis technology and automatic delineation of target area become increasingly mature.However,due to the unfixed shape and size of tumor target,unbalanced tumor image category,low contrast and noise of tumor image,the accuracy of the existing automatic delineation algorithm for tumor target is difficult to meet the requirements of clinical application.In this paper,based on the clinical practice,the automatic segmentation algorithm of tumor target area is studied from the representative tumor diseases in different parts of the human body,such as nasopharyngeal cancer,lung cancer and cervical cancer.In order to solve the problems of image noise,low contrast of tumor target and variable shape and size of tumor target in 3D tumor segmentation,a global multi-level attention network was proposed.The global multi-level attention network integrates the channel and spatial attention mechanism of the visual model as well as the idea of residual structure into each encoder and decoder in the network to strengthen the learning of key areas in tumor images in the network training stage and improve the recognition ability of irregularly shaped target areas.It can effectively reduce the interference of low contrast and noise on model training or reasoning.Experimental results show that this network can better complete the task of 3D tumor target segmentation of multiple diseases.In order to solve the problems of category imbalance and high requirements for hardware equipment in the process of 3D segmentation network training.In this paper,a two-stage tumor target segmentation algorithm based on anomaly perception is proposed.The algorithm fully considers the characteristics of three-dimensional tumor images and transforms the problem of three-dimensional tumor target segmentation into two stages:abnormal slices classification and two-dimensional target segmentation,which is more consistent with the clinical drawing process and enhances the availability,versatility and reliability of the algorithm.Firstly,in the abnormal slices classification stage,a tumor target anomaly perception algorithm based on decision fusion was proposed to solve the problems of small size of training data and poor interpretability of the algorithm.Before the training,the center cutting hybrid algorithm was used for offline expansion of the lung cancer focus training data set.In the training stage,a variety of data enhancement methods were randomly combined to enhance the understanding ability of the tumor target anomaly sensing algorithm for the data set.During the inference process,the decision results of multiple networks were fused by voting.In addition to improving the accuracy of classification,it also enhances the interpretability of classification results.Finally,combined with the characteristics of tumor target image,the post-processing method was used to further improve the classification accuracy.In the stage of two-dimensional tumor target segmentation,a split-attention coding network is proposed.In order to solve the problems of small training data scale and low model generalization performance,the split attention network model pre-trained on ImageNet data set is migrated to the encoder part of the split network to improve the convergence effect of the network on small-scale tumor target data set.From the perspective of network architecture design.The design paradigm of split-attention coding networks is analyzed through detailed experiments.