Research On Dual-modality PET-CT Images Segmentation Of Nasopharyngeal Carcinoma Based On Deep Learning

Nasopharyngeal carcinoma(NPC)is a prevalent malignant tumor,especially in southern China,Southeast Asia.Radiation therapy is the preferred treatment method for NPC.After radiotherapy,the five-year survival rate of NPC patients is more than 50%.In the process of developing a radiotherapy plan,target delineation is a fundamental and crucial step.Currently,target area is delineated manually slice-by-slice by radiologists,which is a time-consuming and tedious task.Therefore,automatic NPC segmentation computer aided system is highly desired in clinical practice because it alleviates clinicians’workload.Computed tomography(CT)image is the necessary image modality to develop a radiotherapy plan.However,nasopharyngeal tumor commonly does not have clearly visible feature in CT image.Nasopharyngeal tumor appears as highlighted area in positron emission tomography(PET)image,but PET image’s spatial resolution is low,and some normal tissues(such as the brain)also appear as highlights in PET image.Thus,only PET image cannot describe accurately tumor’s boundary information.In clinical practice,PET and CT images are often referenced together.PET image provides the tumor’s location information,CT image indicates the normal tissue’s location and boundary.The combination of the two modalities provides a more accurate reference for the target area delineation.In recent years,medical image segmentation based on deep learning has achieved comparable results with human experts.In this work,we utilized U-net network architecture that commonly used in medical image segmentation to segmentation of NPC based on dual-modality PET-CT images.Prior to segmentation,3D affine registration among different cases is implemented to align the images to a uniform space.In order to alleviate the overfitting problem,online data augmentation was carried out during the training.To use the information of two modalities,we assigned two input channels.In order to deal with the problem of severe imbalance of class distribution in medical image segmentation,according to our data characteristics,a weighted cross entropy loss function is designed to increase the loss of tumor pixels.We have also proposed a simple and effective post-processing step to reduce the number of false positive pixels.Threefold cross validation experiments obtained an average Dice score of 0.8614.In order to extract highly expressive features,the network architecture is deep,which may cause the problem of gradient disappear and make the parameters of the lower layers are not trained to be optimal.In response to this problem,we proposed a U-net network with auxiliary paths.The deep supervision introduced by the auxiliary paths explicitly guides the optimization of lower layer’s parameters,enabling these layers to learn more discriminative features.By adding auxiliary paths,the threefold cross validation yielded an average Dice score of 0.8747,which outperforms state-of-the-art methods in NPC segmentation field,and suggests that adding auxiliary paths is a useful strategy.We also evaluated the effects of several main components of the proposed method on the final segmentation performance by threefold cross validation experiments,such as the registration process,post-processing process,and the influence of CT images.In addition,we also explored the potential of different loss functions and different network architecture in our NPC segmentation task.