Convolutional neural network for dynamic pet frame clustering
A dynamic frame reconstruction apparatus and method for medical image processing is disclosed which reduces the computationally expensive reconstruction of images but which retains the accuracy of the image reconstruction. A convolutional neural network is used to cluster the dynamic data into groups of frames, each group sharing similar radiotracer distribution. In one embodiment, groups of frames that have similar reconstruction parameters are determined, and scatter and random estimations are computed once and shared among each of the frames in the same frame group.
1 . An image processing apparatus, comprising:
processing circuitry configured to
receive list-mode data corresponding to a plurality of detection times,
generate a plurality of frames based on the list-mode data,
assign at least one frame of the plurality of frames into a first frame group based on a similarity of each frame within the first frame group,
assign at least one frame of the plurality of frames into a second frame group based on a similarity of each frame within the second frame group,
estimate first and second frame group-specific reconstruction parameters based on (1) frames in the first frame group, and (2) frames in the second frame group, respectively, wherein the first and second frame group-specific reconstruction parameters are different,
reconstruct a first set of frame data from any frame of the first frame group based on the first frame group-specific reconstruction parameters, and
reconstruct a second set of frame data from any frame of the second frame group based on the second frame group-specific reconstruction parameters,
wherein the processing circuitry is further configured to perform clustering using a neural network, and
the processing circuitry is further configured to produce a set of latent features from the plurality of frames the set of latent features being extracted from the neural network, assign the at least one frame of the plurality of frames into the first frame group based on a similarity of latent features of the set of latent features of each frame within the first frame group, and assign the at least one frame of the plurality of frames into the second frame group based on a similarity of latent features of the set of latent features of each frame within the second frame group.
2 . The image processing apparatus according to claim 1 , wherein the processing circuitry is further configured to estimate the first and second frame group-specific reconstruction parameters based on (1) a last-in-time frame in the first frame group, and (2) a last-in-time frame in the second frame group.
3 . The image processing apparatus according to claim 1 , wherein the processing circuitry is further configured to estimate the respective frame group-specific reconstruction parameters from (1) a single frame for a frame group having only one frame, and (2) less than all frames for frame groups having more than one frame.
4 . The image processing apparatus according to claim 1 , wherein the processing circuitry is further configured to estimate the respective frame group-specific reconstruction parameters from (1) a single frame for a frame group having only one frame, and (2) an average of plural frames for frame groups having more than one frame.
5 . The image processing apparatus according to claim 1 , wherein the plurality of frames generated by the processing circuitry is a plurality of frames of crystal counts.
6 . The image processing apparatus according to claim 1 , wherein the first and second frame group-specific reconstruction parameters used by the processing circuitry are first and second frame group-specific scatter parameters.
7 . The image processing apparatus according to claim 1 , wherein the first and second frame group-specific reconstruction parameters used by the processing circuitry are first and second frame group-specific random event parameters.
8 . The image processing apparatus according to claim 1 , wherein the plurality of frames generated by the processing circuitry is a plurality of pre-reconstruction data frames.
9 . The image processing apparatus according to claim 8 , wherein the plurality of pre-reconstruction data frames generated by the processing circuitry is a plurality of crystal count maps.
10 . The image processing apparatus according to claim 8 , wherein the plurality of pre-reconstruction data frames generated by the processing circuitry is a plurality of frames of sinogram data.
11 . The image processing apparatus according to claim 1 , wherein the plurality of frames generated by the processing circuitry is a plurality of preview reconstruction frames without scatter correction.
12 . The image processing apparatus according to claim 1 , wherein the processing circuitry is further configured to perform clustering to assign the at least one frame of the plurality of frames into the first frame group based on the similarity of each frame within the first frame group, and assign the at least one frame of the plurality of frames into the second frame group based on the similarity of each frame within the second frame group.
13 . The image processing apparatus according to claim 1 , wherein the processing circuitry is further configured to:
produce a set of latent features from the plurality of frames,
cluster the set of latent features from the plurality of frames,
assign the at least one frame of the plurality of frames into the first frame group based on the clustered latent features of the set of latent features of each frame within the first frame group, and
assign the at least one frame of the plurality of frames into the second frame group based on the clustered latent features of the set of latent features of each frame within the second frame group.
14 . The image processing apparatus according to claim 1 , wherein
the plurality of frames are temporal frames,
the first frame group and the second frame group are clustered with respect to the temporal frames.
15 . The image processing apparatus according to claim 1 , wherein the processing circuitry is further configured to:
generate a plurality of temporal frames based on the list-mode data,
assign at least one temporal frame of the plurality of temporal frames into a first temporal frame group based on a similarity of each temporal frame within the first temporal frame group,
assign at least one temporal frame of the plurality of temporal frames into a second temporal frame group based on a similarity of each temporal frame within the second temporal frame group,
estimate first and second frame group-specific reconstruction parameters based on (1) temporal frames in the first temporal frame group, and (2) temporal frames in the second temporal frame group, respectively, wherein the first and second frame group-specific reconstruction parameters are different,
reconstruct a first set of temporal frame data from any temporal frame of the first temporal frame group based on the first frame group-specific reconstruction parameters, and
reconstruct a second set of temporal frame data from any temporal frame of the second temporal frame group based on the second frame group-specific reconstruction parameters.
16 . The image processing apparatus of claim 1 , wherein each frame includes frame data being a sinogram or a crystal count map, and the processing circuitry is further configured to assign the at least one frame into the first frame group based on a similarity of the sinogram or crystal count map of each frame within the first frame group.
17 . An image processing method, comprising:
receiving list-mode data corresponding to a plurality of detection times;
generating a plurality of frames based on the list-mode data;
assigning at least one frame of the plurality of frames into a first frame group based on a similarity of each frame within the first frame group;
assigning at least one frame of the plurality of frames into a second frame group based on a similarity of each frame within the second frame group;
estimating first and second frame group-specific reconstruction parameters based on (1) frames in the first frame group, and (2) frames in the second frame group, respectively, wherein the first and second frame group-specific reconstruction parameters are different;
reconstructing a first set of frame data from any frame of the first frame group based on the first frame group-specific reconstruction parameters; and
reconstructing a second set of frame data from any frame of the second frame group based on the second frame group-specific reconstruction parameters,
wherein the method further comprises performing clustering using a neural network, and
the method further comprises producing a set of latent features from the plurality of frames, the set of latent features being extracted from the neural network, assigning the at least one frame of the plurality of frames into the first frame group based on a similarity of latent features of the set of latent features of each frame within the first frame group, and assigning the at least one frame of the plurality of frames into the second frame group based on a similarity of latent features of the set of latent features of each frame within the second frame group.
18 . A non-transitory computer-readable medium storing a program that, when executed by processing circuitry, causes the processing circuitry to perform an image processing method, comprising:
receiving list-mode data corresponding to a plurality of detection times;
generating a plurality of frames based on the list-mode data;
assigning at least one frame of the plurality of frames into a first frame group based on a similarity of each frame within the first frame group;
assigning at least one frame of the plurality of frames into a second frame group based on a similarity of each frame within the second frame group;
estimating first and second frame group-specific reconstruction parameters based on (1) frames in the first frame group, and (2) frames in the second frame group, respectively, wherein the first and second frame group-specific reconstruction parameters are different;
reconstructing a first set of frame data from any frame of the first frame group based on the first frame group-specific reconstruction parameters; and
reconstructing a second set of frame data from any frame of the second frame group based on the second frame group-specific reconstruction parameters,
wherein the image processing method further comprises performing clustering using a neural network, and
the image processing method further comprises producing a set of latent features from the plurality of frames, the set of latent features being extracted from the neural network, assigning the at least one frame of the plurality of frames into the first frame group based on a similarity of latent features of the set of latent features of each frame within the first frame group, and assigning the at least one frame of the plurality of frames into the second frame group based on a similarity of latent features of the set of latent features of each frame within the second frame group.