MR-based attenuation correction for whole-body PET-MR

Objectives In order to assess the potential and quantify the performance of tissue-class segmentation for whole-body (WB) MR-based attenuation correction (AC), we applied 3-, 4-, 5-, and 6-class segmentation directly on CT (CT-based) to obtain the ideal attenuation maps in 23 WB PET-CT patient scans. We then applied the same methods to a MR (MR-based) for 1 patient. We computed the PET bias images.

Methods Tissues were classified into 6 classes: air, lung, fat, soft-tissue, low-density (LD) bone, and high-density (HD)bone. Approaches classifying air, lung, other (3C); air, lung, fat, other (4C-F); air, lung, bone, other (4C-B); air, lung, fat, soft-tissue, bone (5C); air, lung, fat, soft-tissue, LD bone, HD bone (6C); were considered. For the CT-based study, we applied multiple thresholds to the CT for segmentation and obtained the pseudo-CTs. For 1 patient, a MR was acquired with Dixon sequence and registered to the CT. Fat was identified by applying a threshold on the fat images. Other tissue classes were identified using the in-phase MR images. We applied k-mean within a mask, which encloses all the bones, to identify the bones. We reconstructed PET using the clinical protocol. PET bias images were computed using the PET, which was reconstructed with the original CT, as reference.

Results High positive and negative biases were observed in the fat and the bone regions, respectively, with 3C. High negative bias in almost all the bone regions was associated with 4C-F while high positive bias in the fat regions was associated with 4C-B. High negative bias was observed in HD bone regions exclusively for 5C while small bias (<10%) was achieved across the WB with 6C. Our results were confirmed in the MR-based AC in well registered regions. Additional bias was observed in bony structures due to the inaccuracy of bone identification with MR.

Conclusions Our CT-based study shows that achieving accurate (bias<10%) PET quantitation using MR-based AC across the WB(except HD bone regions) requires the use of at least 5C. Fat and in-phase MR can be used to achieve at least 5-class segmentation