医学图像中非刚体配准的gpu优化研究-计算机技术专业论文.docx

配准点，1G规模的3DCT体数据情况下程序配准用时3214秒，CUDA程序同样情况下配准用时70秒，加速比接近46。由于多核等技术可以让图像处理速度线性提升，不需要对源程序作大幅修改，选用最新GPU如英伟达TeslaK80，以及多GPU，多CPU的单台计算机有望达到加速比200，达到实时配准的效果。关键词：薄板样条TPS医疗影像配准解线性方程组矩阵向量乘三次样条插值万方数据ARESEARCHONACCELERATION0FNON-RIGIDREGISTRATIONONMEDlCALlMAGEBYUSlNGGPUHARDWAREABSTRACTRegistrationisakeyfactorinmedicalimageprocessing．AregistrationprocedurecanjointCTorMRIimagestoultrasoundimageinrealtimewhichcouldmakeupthedefectsofeffectofdifferentimagingdevices．Theregistrationcouldalsohelpdoctorstomakeuptheirmindonhumanbonesorthodonticsurgerypreparation．ThispaperproposeamethodofThinPlateSpline(TPS)registrationbyGPUaccelerationonmedicalimage，includingthefollowingparts：theusefulnessofmedicalimageregistration，GPUhardwarelogical，theaccelerationalgorithmofmatrixoperation，thefundamentalprinciplesofTPSmedicalimageregistration，andtheprogrammingofCUDAacceleration．Themainpointofourexperimentisacceleratingmatrixmultiplication．OurexperimentalplatformisINTELXEONE1230V2，8GRAM，GTX460displaycard．Optimizationprocedureincludesacceleratememoryaccess，bankconflict，matrixrecomblnationmbination．，etc．The，1 eelapsedelapsedtimetimeofOIourfirstIlrStCUDAm￡ttrIN,‘multiplicationversionis 0．3286s，compareto the final versionis 0．025s，accelerationratiois13．Thequantityofcontrolpointsisalsoimportanttothecomputation，becausethesizeofthematrixisdependedonthequantityof controlpoints．Theexperimentshowsthattheelapsedtimeof2DpictureIII万方数据registrationisO．O106s、；vith65controlpointsand0．0044swith3controlpoints onMATLAB，whileonVB．net2010elapsedtimeis3214sin1GbyteCTvolumedata．ComparingtoCUDAmethodis70s，theaccelerateratiois46．Inthecaseofusageofmulti—coreprocessorinimageprocessing，theprocessingspeedisnearlylinearincreasing．Withoutmajorchangetothesourcecode，upgradingthehardwaretoNVIDIATeslaK80，ormulti—GPU，multi_CPUsystemcouldrisethespeed—upratioto200，implyingthattherealtimeregistrationisverypromising．Keywords：ThinPlateSpline(TPS)，medicalimageregistration，solvingnonlinearequation，cubicsplineinterpolation，matrix—vectormultiplica