Table 5.

Comparison of the four attenuation correction techniques outlined in this review

	MR-based AC	Emission-based AC	Atlas-based AC	Deep learning-based AC
Accuracy	Low but somewhat improved when properly incorporating bone tissue [36, 73]. Low accuracy in lung and pelvis [120, 241, 308]	Good accuracy in the brain and tissue lesions. Moderate accuracy in air cavities, bone and lung [36, 142]	Good accuracy in the brain for most methods [36]. Moderate in whole-body (limited number of studies) [6]	Good accuracy in brain and body [242, 271]
Artefacts and biases	Truncation artefacts [26]. Motion artefacts in lung and heart [22] Metallic artefacts when implants are present [47]. Workaround techniques to partially alleviate the artefacts [49, 97]	Additive constant [132]. Positive bias on low count data [133]. Crosstalk in non-TOF data [131]	Moderately sensitive to metallic and truncation artefacts [271]. High biases for non-standard anatomies [10, 192]. Separate adult and paediatric databases required [33]	Insensitive to metallic and truncation artefacts [271]. Separate adult and paediatric network training may be required [212]. Insensitive to tracer (depending on the technique) [228, 255]
Processing time	~ 20 s—10 min depending on the sequence [52, 109]	~ 1 h [253]	30 min—several hours [167, 185]	Few seconds—few minutes [242, 271]
Provided by the manufacturers	Yes	No. Only for addressing truncation artefacts on the Siemens mMR [158]	One method that requires T1 images is available on the GE SIGNA [170]	No
User input	Minimal. Only the acquisition of the sequences required	Moderate/High. Increases when coregistration is also required	Moderate/High. Data acquisition, potential data processing and visual inspection	Moderate. After training only the trained weights need to be applied on the new image
Susceptibility to misregistration	Yes [14]	Not for the original MLAA. Yes for methods requiring anatomical priors or initial μ values [137]	Yes. Most methods require two or more registration steps [43, 60]	Yes, if an anatomical MR image is used as input [247]
PET dependency	Independent	Not suitable for non-TOF systems and low count datasets [130]	Independent	Independent if anatomical images are used. Could be tracer-dependent if NAC images are used as input [255]
Applicability to whole-body	Only Dixon-based sequences [67]	Yes [142]	Separate atlases for each region. Not widely used in whole body [186]	Yes [242, 248]
Requirement for additional data	No	Tissue priors required to tackle the additive constant issue	MR images if they are used as input	MR images if they are used as input
Additional requirements	None	Coregistration and/or segmentation tools	Large database of paired CT and MR images Coregistration and/or segmentation tools	Large image database for model training on a powerful workstation or a suitable pre-trained model if available