Improvements in quality and quantification of 3D PET images
The resolution of PET depends on some effects impossible to be readily corrected The implementation of 3D spatially-variant PSF (radially asymmetric) in a 3D iterative reconstruction algorithm has been proposed Simple acquisition scheme and measurement procedure for a 3D mapping of the scanner Some novel elements have been proposed fits accounting for the dimensions of the source transposed PSF in the image space kernel filled with the integral of the PSF Higher contrast recovery, lower noise and more defined volumes Introduction of an edge artifact Increase of noise as iterations proceed (iterative reconstruction algorithm) Proposal of an image regularization scheme Proposal of a new variational prior (dealing with the gradients in the image), having a (tunable) double behaviour: smoothing of background regions edge preservation (with natural appearance) in signal regions Diapositiva 1
The proposed regularization depends on two parameters Proposal of an objective method to optimize these parameters Proposal of the detectability, an image quality inde The optimal set of regularization parameters is the one maimizing the detectability Validation of the detectability inde Maimization of the detectability inde in clinical-like conditions Implementation and use of the regularization scheme with the optimal parameters huge improvement in background noise Quantitative validation: improvement in quantitative content with respect to standard OSEM fast convergence improved uniformity in noisiest regions (i.e. liver) Qualitative validation: increase in contrast and definition of the lesions and activity distribution Diapositiva 2
U ( λ) φ( λ ) Ω = dω λ [ ] = ( ) φ λ λ D λ ( ) > 0 φ ( ) < 0 φ ( ) = 0 φ smoothing the image, lowering the noise level but reducing the sharpness of the edges in the image (with increasing effect for larger values) enhancement of the edges possible amplification of noisy tetures and creation of patchy artefacts (with increasing effect for larger absolute values) neither smooths nor enhances edges ecellent edge preservation and smoothing of flat regions regions with gradual variations: unnatural staircasing effects What is needed in PET with PSF recovery? strong noise suppression (PSF recovery enhances correlation) high signal content preservation (and spatial resolution), but not too much Diapositiva 3
Università degli Studi di Milano-Bicocca XXIV Ciclo di Dottorato in Fisica Improvements in quality and quantification of 3D PET images E. Rapisarda Diapositiva 4 ( ) ( )( ) ( ) + + + < = φ c 3 3 3 4/3 1 ln 1 1 4 3 ( )( ) [ ] 4 ln1 1 1 3 4 / 3 3 3 + + = c ( ) ( )( ) ( ) [ ] + + < = φ 2 3 3 3 3 2 1 1 1 27 1 ( ) ( )( ) ( ) + + < = φ 3 3 3 3 1 1
Gauss-Total Variation φ GTV ( ) 2 2 = 2 2 < φ GTV ( ) = 1 0 < potentially subotpimal smoothing of noise potentially creates unnatural flattening p-gaussian P = p φ [ 1 < p < 2] p 2 φ P (, p) = ( p 1) p strong smoothing at low gradients partial preservation of spatial resolution Modification of the p-gaussian (p=4/3) to improve the preservation of spatial resolution: φ ( ) = d 3 4 / 3 ( d + ) ln 4 + d + c < Diapositiva 5
Diapositiva 6
Diapositiva 7
Both priors depend on two parameters: β (regularization strength) and (background-signal threshold) Optimization by maimizing the detectability inde proposed and validated D µ / µ 1 R 1 S B = 100 ln µ S σ S µ B + σ B Optimization in clinical conditions NEMA IEC Body phantom: fillable tank with fillable spheres ( lesions, inner diameters are [mm] 10, 13, 17, 22, 28 and 37) Lesion-to-background ratio = 4.4 : 1 Time of acquisition: 2 min, total coincidences acquired: 52.4 10 6 Reconstruction up to 50 iterations with PSF and priors, changing β and Analysis of the detectability for the smallest sphere (radius 5 mm) Diapositiva 8
Proposed prior Gauss-Total Variation prior β = 0.002 = 0. 3 β = 0.015 = 0. 2 parameter is the most delicate of the two parameters Along the direction the response slowly changes around the maimum Along the direction the response abruptly changes around the maimum Diapositiva 9
The validation was performed by comparing different reconstructions: standard OSEM (RnoPSF) OSEM with PSF recovery with clinical postfiltering (RwPSF-Filt) standard OSEM with clinical postfiltering (RnoPSF-Filt) OSEM with PSF recovery and Gauss-TV prior (RwPSF-GTVR) OSEM with PSF recovery and the proposed prior (RwPSF-R) Quantitative validation OSEM with PSF recovery (RwPSF) OSEM with PSF recovery and p-gaussian prior (RwPSF-PR) NEMA IEC IQ phantom Fillable tank containing si fillable spheres having different diameters Background act. conc.: 5 kbq/ml Signal:background = 4.4:1 Total prompts = 52.4 10 6 Qualitative validation Two quantitative figures of merit: Background COV STD COV = µ B = 1 1 µ N voels 1 Contrast recovery CR hot ( Ci µ B ) i V Oncologic patients (5 iterations) B µ S µ B 1 = 100 R 1 2 Diapositiva 10
Three uniform spheres on FOV=700 mm, 256 piels 256 piels, 28 subsets, 10 iters Effect linked with the recovery of PSF Two-dimensional simulation: reconstruction of a circle superimposed to a given background Correct PSF 280 iterations Wrong PSF 280 iterations Correct PSF 28000 iterations Wrong PSF 28000 iterations Diapositiva 11
Comparison of the different reconstruction algorithms (28 subs., 5 it., 256256) Diapositiva 12
Diapositiva 13
Diapositiva 14
FOV=600 mm, 256 piels 256 piels, 28 subsets, 10 iters RnoPSF RnoPSF-Filt RwPSF RwPSF-Filt RwPSF-GTVR RwPSF-PR RwPSF-R Diapositiva 15
Diapositiva 16
Diapositiva 17
Diapositiva 18
Diapositiva 19
With respect to RwPSF-Filt Diapositiva 20
5 iterations, FOV 60 cm RwPSF RwPSF-Filt RwPSF-R Diapositiva 21
5 iterations, FOV 60 cm RwPSF-GTVR RwPSF-PR RwPSF-R Diapositiva 22
Installation finished on the end of Oct. 2010 Acceptance and NEMA NU-2-2007 performance tests performed FWHM (mm) Radial and Tangential Average 4.70 Measurement and implementation of PSF Implementation and optimization of regularization Spatial Resolution (mm) Sensitivity (cps/kbq) 0 cm 10 cm FWTM (mm) FWHM (mm) FWTM (mm) Aial 4.74 Radial and Tangential Average 8.83 Aial 10.91 Radial 5.34 Tangential 4.79 Aial 5.55 Radial 10.07 Tangential 8.96 Aial 11.14 0 cm 7.4 10 cm 7.6 Qualitative and quantitative comparison between different reconstruction algorithms Scatter Fraction (%) 37 Maimum absolute error at NECR peak (%) 2.09 Peak NECR ( kcps, kbq/ml) 139.1 29.0 Diapositiva 23
Diapositiva 24
Diapositiva 25
β = 0.003 = 0. 4 Diapositiva 26
The validation was performed by comparing different reconstructions: standard OSEM (RnoPSF) standard OSEM with clinical postfiltering (RnoPSF-Filt) OSEM with PSF recovery (RwPSF) OSEM with PSF recovery with clinical postfiltering (RwPSF-Filt) TOF OSEM (TOF RnoPSF) TOF OSEM with clinical postfiltering (TOF RnoPSF-Filt) TOF OSEM with PSF recovery (TOF RwPSF) TOF OSEM with PSF recovery with clinical postfiltering (TOF RwPSF-Filt) TOF OSEM with PSF recovery and p- Gaussian prior (TOF RwPSF-PR) TOF OSEM with PSF recovery and the proposed prior (TOF RwPSF-R) Quantitative validation NEMA IEC IQ phantom Signal:background = 4.4:1 Total prompts = 50.2 10 6 Background COV Hot contrast recovery Cold contrast recovery COV = STD 1 1 2 µ S µ B 1 = ( Ci µ B ) CRhot = 100 µ µ N 1 R 1 B B voels i V CR cold µ C 100 1 µ = B Qualitative validation Oncologic patients (10 iterations) Diapositiva 27
FOV=600 mm, 256 piels 256 piels, 18 subsets, 10 iters Non TOF RnoPSF RnoPSF-Filt RwPSF TOF Diapositiva 28
Diapositiva 29
Diapositiva 30
Diapositiva 31
With respect to RnoPSF With respect to TOF RnoPSF Diapositiva 32
FOV=600 mm, 256 piels 256 piels, 18 subsets, 10 iters TOF RwPSF TOF RwPSF-Filt TOF RwPSF-PR TOF RwPSF-R Diapositiva 33
Diapositiva 34
Diapositiva 35
Diapositiva 36
Diapositiva 37
With respect to TOF RnoPSF With respect to TOF RwPSF-Filt R vs. PR 5 it 10 it 50 it Cold, large +0.1% +0.1% +0.5% Hot, small +11.2% +16.3% +17.8% Noise +1.0% +2.1% +3.2% Diapositiva 38
FOV=600 mm, 256 piels 256 piels, 18 subsets, 10 iters TOF RwPSF-Filt TOF RwPSF-PR TOF RwPSF-R Diapositiva 39
FOV=600 mm, 256 piels 256 piels, 18 subsets, 10 iters TOF RwPSF TOF RwPSF-Filt TOF RwPSF-PR TOF RwPSF-R Diapositiva 40
The use of PSF recovery inside an iterative reconstruction algorithm allows improving the quality and the quantitative content of PET images Introduction of an edge artifact Increase of noise as iterations proceed (iterative reconstruction algorithm) Proposal of an image regularization scheme The proposed prior, compared to the chosen competitors, allows controlling the edge artefact without significant loss of spatial resolution Convergence process speeded up allows controlling the noise increase while retaining good quantitative performance Safer prior than Gauss-TV Direct comparison with p-gaussian: drawbacks from the proposed strategy are much less important than benefits Confirmation of the results also if the TOF information is taken into account, in particular in the coldest regions Diapositiva 41
I can no other answer make, but thanks, and thanks. W. Shakespeare Thanks to all the people who supported me, in any way, at any time. And thank you for not sleeping. Diapositiva 42