Spinal Neoplasm Image Inpainting Based On Deep Learning

Spinal tumor is an abnormal mass of tissue in or surrounding the spinal cord or spinal column that often grows rapidly,leading to paralysis and other complications,seriously affecting the quality of life of patients.Total En Bloc Spondylectomy is an effective treatment which can remove tumor lesion with tumor-free surgical margin.Owing to the removal of posterior vertebral elements and anterior parts,the continuity of the spine is interrupted and the stabilization of the spine is considerably necessary.At present,3D printing has a great help for the treatment of the difficult cases and surgery which can print a 1:1 3D spinal model and achieve individual customization.However,there are some limitations in the application of 3D printing to clinical now.It is difficult to completely segment the region of cancellous bone and soft tissue in CT images,so that the prosthesis produced will not fit perfectly with the patient.What’s more,without sufficient medical knowledge,it is likely that the prosthesis produced by the revised 3D digital model may differ from the ideal one.Therefore,making multiple pre-operative models for discussion is often necessary which result in the increase of the extra time and overall cost.To solve the above problems,this paper proposes to carry out semantic segmentation,lesion detection and image inpainting through deep learning to reconstruct a normal and complete vertebral body for 3D printing.In this way,solve the problems of 3D printed spinal prosthesis,promote the further application of 3D printing in the field of spinal surgery,and assist doctors improve the accuracy of tumor detection.Firstly,the images of spine CT are segmented by deep neural network in this paper to obtain binary images with the spinal region as the foreground and the rest region as the background,providing the basis for subsequent lesion detection and image inpainting.Then,this paper is based on the deep convolution network to extract the feature of the spinal tumor image,obtain the high level of the data abstract representation and locate the lesions of the spine.In order to reconstruct the complete spinal 3D model,the generative adversarial network is applied for the image inpainting of the spinal tumor CT images in this paper.The network is trained using binary images of normal spinal CT to obtain the corresponding model firstly.Then,add mask to the spinal lesion region as the input of the model to repair the images and the output of network is restored CT images.Finally,through 3D reconstruction,a complete three-dimensional model of the spine is obtained.The spinal CT images are segmented through the deep convolution neural network in this paper which has a good robustness,achieves more than 95% accuracy and avoids the effect of the imprecise edge segmentation of the traditional image segmentation method.For the task of detection and classification of spinal tumors,the detection and identification of lesions in the image of spinal tumors by deep convolution neural network can be achieved by using data pre-processing and labels,so as to accurate detection of tumor areas.The lesion regions of spinal CT images are repaired by the generative adversarial network to obtain the complete spinal images and achieve the reconstruction of complete spinal model in 3D,filling the gap in this aspect and providing the basis for the production of spinal prosthesis.