Research On Segmentation Of Organs At Risk In Thoracic Image Based On Deep Learning

Currently,radiation therapy is one of the relatively effective means of treating cancer by using α、β、γ and various types of X-rays to kill DNA of cancer cells which can stop the abnormal division and proliferation of cancer cells,so as to accomplish the diagnosis and therapy of cancer.The cancerous area and nearby healthy organs named organs at risk(OARs),the most important step in the whole treatment plan is to finely outline the contours of OARs and their boundary.Not only manual segmentation is inefficient,but also requires a good knowledge of human anatomy and clinical experience.Due to the complexity of human body structure,the same doctor may have certain deviations when outlining the same CT image at different time,which will make it difficult to ensure the quality of radiation treatment.The segmentation algorithm based on deep learning can achieve end-to-end accurate segmentation of organs at risk and assist doctors in completing radiotherapy with high quality and efficiency which has a certain contribution to the development of clinical medicine.Starting from the characteristics of organs at risk in thoracic,this paper focuses on multi-scale featureaware encoding-decoding network and the segmentation network model based on multi-scale attention,and propose two solutions based on deep learning.The main work is as follows:(1)In this paper,a multi-scale feature-aware encoding-decoding network(FAUnet)is proposed to segment OARs in thoracic CT image.To address the problem of the difference in size between the four organs in thoracic,we first design an inputaware module to extract multi-scale features of the target object in both large and small organs.In order to bridge the semantic gap between the encoding and decoding layers,the modified Inception module is introduced to long-range skip connections between the encoding part and the decoding part in our architecture;Furthermore,we leverage the Efficient Spatial Pyramid(ESP)and Pyramid Spatial Pooling(PSP)module replace the traditional serial convolutional operations,making the network more lightweight and alleviating the overfitting matter create by scarcing data.To settle the issue of class imbalance,we formulate a novel loss function by combining Dice coefficient and Cross Entropy to train our network.The experimental indicate that the network we propose has achieved competitive advantages in accuracy and computational efficiency,and the segmentation of multi-target OARs with different shapes and sizes such as the heart,aorta,esophagus,and trachea is more refined.(2)Aiming at the above-mentioned FA-Unet’s inaccurate segmentation of the isolated part in thoracic CT image and the unsatisfactory effect of boundary segmentation,we propose a multi-scale attention-based network model for OARs in thoracic segmentation(Multi-Attention U-Net,MA-Unet).Firstly,a multi-scale attention is designed to combine the multi-scale output of the decoded module with the channel attention module to improve the segmentation of tiny isolated parts in the thoracic.Secondly,we make the output feature vector of multi-scale attention elementwised with the shallow feature map,using Deep Feature-Guided Shallow Feature(DFGSF)which effectively combines the contextual information between the deep and shallow layers in different level to improve the segmentation effect of the network on the boundary of the organs.Experimental results show that,compared to FA-Unet and other U-shaped encoding and decoding network models of the same type,the MA-Unet proposed in this paper has a better segmentation effect on various OARs in thoracic image.Especially for organ edges and tiny isolated parts,the segmentation is more refined.