Hybrid approach for attenuation correction in PET/MR scanners

Aim: Attenuation correction (AC) of PET images is still one of the major limitations of hybrid PET/MR scanners. Different methods have been proposed to obtain the AC map from morphological MR images. Although, segmentation methods normally fail to differentiate air and bone regions, while template or atlas methods usually cannot accurately represent regions anatomically different from the template image. In this study a feed forward neural network (FFNN) algorithm is presented which directly outputs the attenuation coefficients by non-linear regression of the images acquired with an ultrashort echo time (UTE) sequence guided by the template-based AC map (TAC-map). Materials and methods: MR as well as CT data were acquired in four subjects. The UTE images and the TAC-map were the inputs of the presented FFNN algorithm for training as well as classification. The resulting attenuation maps were compared with CT-based, PNN-based and TAC maps. All the AC maps were used to reconstruct the PET emission data which were then compared for the different methods. Results: For each subject dice coefficients D were calculated between each method and the respective CT-based AC maps. The resulting Ds show higher values for all FFNN-based tissues comparatively to both TAC-based and PNN-based methods, particularly for bone tissue (D=0.77, D=0.51 and D=0.71, respectively). The AC-corrected PET images with the FFNN-based map show an overall lower relative difference (RD=3.90%) than those AC-corrected with the PNN-based (RD=4.44%) or template-based (RD=4.43%) methods. Conclusion: Our results show that an enhancement of current methods can be performed by combining both information of new MR image sequence techniques and general information provided from template techniques. Nevertheless, the number of tested subjects is statistically low and current analysis for a larger dataset is being carried out.