Metal Artifacts Reduction On Cervical CT Image By Deep Learning

Cervical cancer is the most common gynecological malignancy and the third most common cancer cause of death among women worldwide,Brachytherapy is one of the most effective treatments for cervical cancer.Before brachytherapy,afterloading devices with different types of metallic applicators are generally implanted inside the vaginal cavity of the patients with cervical cancer.During brachytherapy,computed tomography(CT)imaging is necessary since it conveys tissue density information which can be used for dose planning.However,these metallic implants have much higher attenuation coefficients than bones or soft tissue.The X-rays were heavily attenuated after passing through metal objects,resulting in only weak signals reaching the detector.In this situation,if the X-ray detector lacks a sufficient dynamic range in detecting the weak signal,there will be metal shadows in the raw projection data.These metal shadows will introduce streak artifacts,which can spread to nearby soft tissue regions in the reconstructed cervical CT images,obscuring the crucial diagnostic information of the tissues surrounding the metallic implants.These metal artifacts caused by brachytherapy applicators remain a challenge for the automatic processing of image data for image-guided procedures or accurate dose calculation.Therefore,developing an effective metal artifact reduction(MAR)algorithm in cervical CT images is of high demand.In this paper,after analysis and summary of relevant research results about metal artifacts reduction in CT images,the following major works have been done according to metal artifact reduction of cervical CT image based on deep learning.(1)Clinical cervical CT image metal artifact data simulation.Due to the lack of public data set of cervical CT images with metal artifacts and paired cervical CT images without metal artifacts,we propose to simulate metal artifact data of clinical cervical CT images.Under the guidance of oncologists,we manually simulated the shape,size and location of the metallic applicator.Through numerical simulation technology,we simulated striated metal artifacts on the cervical CT images without metal artifacts and obtained the paired cervical CT images with metal artifacts.(2)A simple convolutional neural network based metal artifact reduction algorithm for cervical CT image has been implemented.Using the simulated data set above,cervical CT images with metal artifacts and paired cervical CT images without metal artifacts(as a reference label)were input into the designed convolutional neural network for training,and then the de-metal-artifacts cervical CT images were obtained.Experimental results show that convolutional neural networks can effectively remove metal artifacts in CT images.However,this kind of network output operation leads to residual artifacts in the de-metal-artifact cervical CT image..(3)An improved MAR algorithm for cervical CT images based on residual learning(RL-ARCNN)has been designed,implemented and validated.According to the concept of residual learning,a model for extracting metal artifacts of cervical CT images was constructed and trained with simulation data sets.The input of the model is a cervical CT image with metal artifacts,and the output is the extracted metal artifacts.Finally,the de-metal-artifacts cervical CT images were obtained by the extracted metal artifacts subtracted from the cervical CT images with metal artifacts.The experimental results show that the proposed RL-ARCNN outperforms current conventional MAR methods with respect to metal artifacts reduction and tissue texture preservation.