PET/MR is an emerging hybrid imaging modality. However, attenuation correction (AC) remains challenging for hybrid PET/MR in generating accurate PET images. Segmentation-based methods on special MR sequences are most widely recommended by vendors. However, their accuracy is usually not high. Individual refinement of available certified attenuation maps may be helpful for further clinical applications. In this study, we proposed a multi-resolution regional learning (MRRL) scheme to utilize the internal consistency of the patient data. The anatomical and AC MR sequences of the same subject were employed to guide the refinement of the provided AC maps. The developed algorithm was tested on 9 patients scanned consecutively with PET/MR and PET/CT (7 [18F]FDG and 2 [18F]FET). The preliminary results showed that MRRL can improve the accuracy of segmented attenuation maps and consequently the accuracy of PET reconstructions.
Keywords: Attenuation correction; Machine learning; Multi-resolution; PET/MR.
Copyright © 2016 Elsevier Ltd. All rights reserved.