Research Of Postoperative Gliomas Segmentation On CT Image Based On Deep Learning

Gliomas are the most common tumors in the central nervous system,which has a high fatality rate and seriously threatens human health.Surgery is the primary treatment for gliomas.Since gliomas often invades surrounding normal tissues and organs,to consider protect the normal functions of the human body,it is often difficult for surgical resection to completely remove the tumors.Radiotherapy has become one of the complementary treatments after operation.It can deliver precise radiation doses to a tumor volume while minimizing the dose to nearby healthy tissues,thus improving the local control rate and survival rate of patients.During radiotherapy,the accurate segmentation of tumor is an essential step.The manual delineation of postoperative gliomas is a routine strategy in clinical practice.However,this method is time consuming,tedious and highly reliant on the experience of clinicians.Therefore,we hope to use image segmentation technology to help doctors delineate target areas.The size and shape of gliomas vary among different individuals,and the appearance and location of gliomas vary greatly.In addition,the anatomical structures of gliomas in pre-operative and post-operative are also quite different.The contrast of soft tissue on CT images was low,and the boundary between glioma and surrounding tissues was blurred in post operation.Currently,segmentation of post-operation gliomas on CT images is still an arduous task for radiotherapists.Compared with CT images,magnetic resonance imaging(MRI)images of post-operative gliomas have good soft tissue contrast,especially the characteristics of different modes of MRI images are different,which can provide complementary information about tumors and surrounding tissues.At present,clinicians often delineate tumor volume on CT images of post-operative glioma refer to multi-modal MRI images.Correspondingly,aiming at the problem of post-operative glioma CT image segmentation,this paper studies and achieve a high accuracy automatic tumor segmentation method of the post-operative glioma on CT images guided by the complementary information of multi-modal MRI images based on deep learning in order to help doctors delineate target areas and reduce the workload.Firstly,this paper studies the glioma segmentation of multi-mode MRI image guided CT image under the deep learning framework.We put forward a fully automated Deep-learning-based Multi-source(CT and multi-modality MR images)Integration Model(M2CT_DSM)for postoperative CT image tumor segmentation.In M2CT DSM,the CT and multi-modality MRI images were concatenated into multi-channel images to enhance the contrast of soft tissue in the image space(gray space).Then the multi-channel images and the manual segmentation label were fed into the designed deep segmentation network for training and then the segmentation of the CT tumor region was completed.Experimental results show that M2CT DSM segmentation results are better than those of the deep segmentation model trained by CT images alone.Secondly,we put forward a deep-feature-fusion-model(DFFM)guided by multi-sequence MRIs for postoperative glioma segmentation in CT images.Since the superposition of gray level on the image space will interfere with the feature extraction of different mode images,in order to make better use of multi-mode MRI images to guide postoperative glioma CT image segmentation,our DFFM only combines the high-level feature maps to fuse the CT and MR information.This kind of fusion can preserve the respective information of each task and improve the performance well.Experimental results show that compared with M2CT_DSM,DFFM can obtain more accurate and robust segmentation results.