MR And TRUS Image Denoising And Segmentation Methods In Prostate Puncture Guidance

In recent years,with the deterioration of the environment,the incidence of prostate cancer is increasing,and the importance of early diagnosis of prostate is becoming more and more prominent.Prostate biopsy assisted by Transrectal Ultrasonography(TRUS)and Magnetic Resonance Imaging(MRI)have become important criteria for the diagnosis of early prostate cancer.MR image registration to real-time TRUS image can improve the puncture accuracy in image-assisted prostate puncture surgery.Since the TRUS image and the MR image are contaminated by noise during the imaging process,the image needs to be de-noised.In order to accurately register the MR image to the TRUS image,gland contour need to be segmented.In order to achieve precise puncture of TRUS and MR image-assisted prostate puncture,this paper studies image de-noising and gland contour segmentation methods for prostate TRUS and MR images.This paper analyzed the existing de-noising methods and the characteristics of prostate medical images.and the images are enhanced to preprocess.The anisotropic diffusion de-noising algorithm is combined with the absolute median difference to denoise the prostate image.The image was evaluated by the peak signal-to-noise ratio after de-noising,and the edge retention characteristics of the gland were evaluated by the edge retention.The superiority of the anisotropic diffusion de-noising method combined with absolute median difference improvement in medical image de-noising is verified.Based on the U-net network structure of convolutional neural network,the U-net network segmentation method is improved by using affine transformation and elastic distortion to expand the U-net network training set samples.Improve the stability of the model.The results of test images segmentation and the traditional segmentation method are compared with the "gold standard" image respectively,and the segmentation method proposed is demonstrated to be superior.Combined with the characteristics of TRUS image,the disadvantages of TRUS image training segmentation network through U-net structure are analyzed.The segmentation method of the U-net network improved by combining the residual network is proposed to solve the gradient disappearance problem in the network deepening.The new method speeds up the convergence and improves the segmentation precision.After training by training set,the U-net network improved by the residual network was used to segment the prostate The segmented TRUS images were evaluated by various criteria,and the superiority of the proposed network architecture in image segmentation was verified.In order to verify the feasibility of the image de-noising method and segmentation methods proposed in this paper,an experimental platform was built to carry out the verification experiment.Firstly,the experimental platform was established to complete the MR and TRUS image acquisition and enhancement of the prostate puncture phantom.After the anisotropic diffusion method improved by the absolute median difference,the TRUS and MR images were de-noised and expanded by affine transformation and elastic distortion.The improved U-net network architecture of the training set segmentation of the MR image,and the TRUS image segmentation is performed by combining the improved U-net network architecture with the residual network short connection.Three-dimensional reconstruction and model registration of the segmented TRUS and MR images were proposed.The internal and external parameters of the ultrasound probe were calibrated to obtain the transformation relationship between the intra-operative TRUS image lesion point,the preoperative MR image lesion point and the puncture needle end point.Drive the puncture mechanism to puncture.The puncture experiment was evaluated by the puncture error.Verify the effectiveness of the proposed de-noising and segmentation methods.