Fig 1.
Overflow of the dynamic PET image reconstruction framework.
Fig 2.
Left: Auto-encoder template. Right: SAE model. An autoencoder is a three layer network including an encoder and a decoder. The SAE model is combined by several encoders and a decoder. The hidden layer of an encoder is the input of its next encoder.
Fig 3.
Visualization of filters learned with SAE.
(a) (b) Different physical phantoms. (d) (e) Different features learned from above phantoms. (c) Brain phantoms. (f) Features learned from brain phantoms.
Fig 4.
Restricted Boltzmann machine.
Fig 5.
Reconstruction results of the brain phantom (left) and Zubal phantom (right) for different size of patches.
Fig 6.
Reconstruction results of the brain phantom (left) and Zubal phantom (right) for different numbers of nodes in the hidden layers.
Fig 7.
Monte Carlo simulation results.
Monte Carlo simulation brain phantom data (left) and Zubal phantom data (right).
Fig 8.
Brain phantom reconstruction results.
From top to bottom: ground truth, reconstruction result by MLEM, MLEM+SAE and TV. From left to right: the 1st, 3rd, 5th, 7th, and 9th frames.
Fig 9.
Zubal phantom reconstruction results.
From top to bottom: ground truth, reconstruction result by MLEM, MLEM+SAE and TV. From left to right: the 1st, 3rd, 5th, 7th, and 9th frames.
Fig 10.
Brain phantom reconstruction results for the local patch.
First row: ground truth for the 3rd frame, ground truth for the local patch. Second row: reconstruction result for local patch by MLEM, reconstruction result for local patch by MLEM+SAE.
Fig 11.
Zubal phantom reconstruction results for the local patch.
First row: ground truth for the 3rd frame, ground truth for the local patch. Second row: reconstruction result for local patch by MLEM, reconstruction result for the local patch by MLEM+SAE.
Fig 12.
Brain phantom and Zubal phantom reconstruction result.
Brain phantom (left) and Zubal phantom (right) reconstruction result comparison for different regions of interest. From top to bottom: SNR, bias and variance comparison curves.
Fig 13.
Real heart data reconstructed results.
From left to right: Reconstruction result by the MLEM algorithm for the 2nd, 3rd and 4th frames, our result for the 2nd, 3rd and 4th frames.
Fig 14.
Reconstruction results under different counting rate settings.
From left to right: the counting rates are 5 × 104, 1 × 105, 5 × 105 and 1 × 106. Top row: reconstruction results by MLEM. Second row: reconstruction results by MLEM+SAE. (a) Brain phantom. (b) Zubal phantom.
Table 1.
Brain phantom reconstruction results comparison with different counting rates.
Table 2.
Zubal phantom reconstruction results comparison with different counting rates.
Fig 15.
Reconstruction results for the Zubal phantom data.
Reconstruction results for the Zubal phantom data using the MLEM algorithm (top row) and MLEM+SAE (second row) with a brain phantom for training. From left to right: the 1st, 3rd, 5th, 7th, and 9th frames.
Table 3.
Zubal phantom reconstruction results comparison.