Abstract:PET/MR hybrid imaging systems have been exhibiting significance in clinic trials and diagnosis, such as the diagnosis of central nervous system diseases as well as heart and macrovascular diseases. However, feasible pathways for PET attenuation correction (AC) in terms of MR imaging still belong to ongoing topics. A few methods were proposed to deal with the AC issue on certain body sections, such as brain and pelvis. Owing to multiple factors, e.g., different organ properties, human respiration, organ motion etc., the generalizability of existing methods are usually unsatisfactory. In answer to such challenge, by jointly using transfer fuzzy clustering (TFC) as well as back propagation neural network (BPNN), we design the novel TFC-BPNN-leveraged tissue segmentation based AC method (TFC-NN-TSAC for short) for PET/MR images. Benefitting from historic knowledge reference, TFC is able to initialize the MR data into several categories, e.g., fat, air, bone, and soft tissue. BPNN is used to further identify bone from air. As such, we preferably solve the problem of imperfect information hidden in MR data potentially caused by noise or image artifacts. With the voxel labels obtained by TFC-NN-TSAC, a fixed linear attenuation coefficient of each tissue type is assigned to implement eventual PET attenuation correction. The experimental results demonstrate that the proposed TFC-NN-TSAC method has good performance of tissue segmentation for Dixon-based MR data, thereby owning nice AC performance for PET images.

A Transfer Fuzzy Clustering and Neural Network Based Tissue Segmentation Method During PET/MR Attenuation Correction