Research On Deep Learning Algorithm For Detection And Segmentation Of Nasopharyngeal Carcinoma

Nasopharyngeal carcinoma(NPC)is a malignant tumor with high mortality,which is prevalent in southern China.NPC is more common in Guangdong and Guangxi of China.Since morbidity of NPC ranks first in otolaryngology malignant tumors,reliable automatic segmentation of NPC in medical images is of great significance for the formulation of treatment plans and subsequent treatment evaluation.Magnetic resonance imaging(MR images)is frequently utilized to assess the location and shape of NPC because of its high resolution for soft tissues.However,the shape of NPC is variable and the volume of NPC is usually small,which makes the automatic segmentation of NPC become a difficult task.In order to address this challenge,this paper focuses on the detection and segmentation of NPC within the framework of deep learning.The contributions of this thesis are summarized as follows:(1)This paper proposes an NPC detection model based on deep reinforcement learning.Based on the analysis of the problems in the existing similarity metrics of bounding boxes,a reward function is proposed to provide more reasonable rewards on the actions taken by the agent.In order to reduce the search range of the agent’s exploration,this paper also proposes an exploration strategy based on prior knowledge,which can also accelerate the convergence of training and improve the performance of the model.(2)Based on the existing attention mechanism for deep learning,this paper proposes a recurrent attention mechanism consisting of two recurrent attention modules,which focuses on emphasizing the key channels and regions,respectively.The recurrent attention mechanism utilizes the semantic information of the deeper features to guide the shallower features in a convolutional neural network,emphasizing the features and regions that are highly correlated with NPC to improve the ability of model representation.A new network architecture named Recurrent Attention Network(RANet)is then proposed based on the recurrent attention mechanism.In addition,considering the problem that pixels near the border of NPC in the MR image are difficult to segment precisely,this paper performs the dilation operation to the boundary of NPC,and then pays more attention to pixels of the dilated border of NPC in the loss function,which can lead to a better segmentation of NPC.(3)This paper also combines the detection model mentioned in(1)and the segmentation model(RANet)in(2)to form an integrated model for NPC detection and segmentation,which is based on deep reinforcement learning and recurrent attention mechanism.In summary,this paper proposes a detection model based on deep reinforcement learning,a segmentation model based on recurrent attention mechanism,and a combination of both.This paper evaluates the effectiveness of the proposed methods on 1804 slices of MR images from 596 patients,which are acquired from Sun Yat-sen University Cancer Center.Experimental results show that recurrent attention mechanisms improve the baseline model significantly by increasing a Dice score of 5.60%,and the Dice score of the integrated model combined segmentation and detection achieved 80.32%.Thus,the high performance of the proposed methods is further demonstrated by the experimental results.