US 12,248,045 B2
Collocated PET and MRI attenuation map estimation for RF coils attenuation correction via machine learning
Deepak Bharkhada, Knoxville, TN (US); and Vladimir Panin, Knoxville, TN (US)
Assigned to Siemens Medical Solutions USA, Inc., Malvern, PA (US)
Filed by Siemens Medical Solutions USA, Inc., Malvern, PA (US)
Filed on Sep. 25, 2020, as Appl. No. 17/031,974.
Prior Publication US 2022/0099770 A1, Mar. 31, 2022
Int. Cl. G01R 33/48 (2006.01); A61B 5/055 (2006.01); A61B 6/00 (2024.01); G01R 33/56 (2006.01); G06T 7/00 (2017.01)
CPC G01R 33/481 (2013.01) [A61B 5/055 (2013.01); A61B 6/5247 (2013.01); G01R 33/5608 (2013.01); G06T 7/0012 (2013.01); G06T 2207/10104 (2013.01); G06T 2207/20081 (2013.01)] 19 Claims
OG exemplary drawing
 
1. A computer-implemented method for attenuation correction, comprising:
receiving, by at least one processor, positron emission tomography (PET) time-of-flight (TOF) data generated by a PET imaging modality collocated with a magnetic resonance (MR) imaging modality;
extracting, by the at least one processor, Radio Frequency (RF) coil attenuation data from the PET TOF data, the RF coil attenuation data characterizing an attenuation caused by collocated RF coils of the MR imaging modality;
generating, by the at least one processor, an initial Radio Frequency (RF) coil attenuation map based on the RF coil attenuation data;
applying, by the at least one processor, a trained model configured to improve a signal to noise ratio of the initial RF coil attenuation map to generate a final RF coil attenuation map, the final RF coil attenuation map having a higher signal to noise ratio than the initial RF coil attenuation map and configured to correct for the attenuation caused by the RF coils during acquisition of the PET TOF data, wherein the trained model is generated based on training an untrained model using an iterative training process that comprises:
training a first set of embedding layers of an untrained model based on a set of training data to generate a first RF coil attenuation map; and
subsequent to training the first set of embedding layers, training a second set of embedding layers of the untrained model based on the set of training data and the first RF coil attenuation map generated by the first set of embedding layers, wherein the training of the second set of embedding layers causes a refinement of the first RF coil attenuation map, the training comprising comparing the refined first RF coil attenuation map to one or more ground truth attenuation maps and adjusting at least one of the second set of embedding layers of the untrained model based on the comparison after each iteration;
performing, by the at least one processor, attenuation correction of the PET TOF data based in part on the final RF coil attenuation map; and
reconstructing, by the at least one processor, an image from the attenuation corrected PET TOF data wherein each of the one or more ground truth attenuation maps are generated from PET TOF data generated using a known emission source.