基于扩散循环一致性生成对抗网络的骨盆活跃骨髓区域分割方法

打开文本图片集

Abstract：ObjectiveToestablishapelvicactivebonemarrow（ABM）segmentationmethodbasedondifusioncycle-consistent generativeadversarial networksforimprovingindividualizedprecisionofconventionalanatomicalatlas-basedmethods. MethodsWecolectedpelvicPET-CTdatafrom253patientsandconstructeda3-stagecascadedcross-modal learning frameworkforpreciseindividualizedABMidentificationfromCTimages.Theframeworkusedcycle-consistentgenerative adversarial networks for bidirectional CT-PET mapping，conditional diffsion modules with100-step Markovchains for progresivedenoising，and multi-scale progressve featurepyramid fusionnetworks for segmentation.The peak signal-to noiseratio ，suctualslarityex （S），noaledanareerrr （NE），icelarityefit andaverage symmetricsurface distance （ASSD）wereused for evaluationofthemodelperformance for ABMsegmentation. Results The proposed method outperformed the existingmethodswitha PSNRof 26.42±0.63 dB，an SSI of and anNMSEof00235±.006.ForABMsegmentation，theaverageDicecoeficientof themodelreached0.770023withan ASSD of 3.52±0.41mm .Conclusion Compared with the conventional methods，the propose method significantly improves individualized segmentationaccuracyoftheABMand isthussuitableuseinindividualizedbonemarrowprotection radiotherapy for rectal cancer.

Keywords：rectalanceractivebonemarow;difusionmodels;gnerativeadversarialetworks;imagesegmetatio

局部进展期直肠癌是消化系统常见恶性肿瘤，全球每年新发约80万例，其中 50% 患者初诊时已属局部晚期。（剩余19347字）