Decomposition-based harmonization for quantitative PET imaging across scanners and radiotracers.
Authors
Affiliations (4)
Affiliations (4)
- Beijing Engineering Research Center of Radiographic Techniques and Equipment, Institute of High Energy Physics, Chinese Academy of Sciences, Beijing, China.
- School of Nuclear Science and Technology, University of Chinese Academy of Sciences, Beijing, China.
- Department of Nuclear Medicine/PET Center, Huashan Hospital, Fudan University, Shanghai, China.
- School of Chemical Engineering, University of Chinese Academy of Sciences, Beijing, China.
Abstract
Quantitative positron emission tomography (PET) is widely applied in oncology, neuroscience, and clinical practice. However, its quantitative accuracy is often compromised by systematic variability arising from differences in scanners, acquisition protocols, and radiotracers, which limits the reliability of multicenter studies. To develop and validate PETHarmony, a novel voxel-level harmonization framework for minimizing inter-scanner and inter-tracer variability. PETHarmony utilizes a linear neural network to model covariates and singular value decomposition to isolate and remove variability in voxel space. Its performance was assessed in four scenarios: (i) using paired <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>18</mn></msup> <mi>F</mi></mrow> <annotation>$^{18}{\rm F}$</annotation></semantics> </math> -FBB PET/CT and PET/MR scans (N = 25, Huashan Hospital) to test the removal of scanner variability; (ii) using 20 repeated PET acquisitions of a NEMA NU-2 IQ phantom to validate absolute quantitative accuracy; (iii) using paired <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>18</mn></msup> <mi>F</mi></mrow> <annotation>$^{18}{\rm F}$</annotation></semantics> </math> -FBP and <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>11</mn></msup> <mi>C</mi></mrow> <annotation>$^{11}{\rm C}$</annotation></semantics> </math> -PiB scans from GAAIN (N = 46); and OASIS (N = 84) to evaluate cross-tracer consistency of cortical SUVR; and (iv) using unpaired multicenter data from ADNI (N = 471; <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>18</mn></msup> <mi>F</mi></mrow> <annotation>$^{18}{\rm F}$</annotation></semantics> </math> -FBP, <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>18</mn></msup> <mi>F</mi></mrow> <annotation>$^{18}{\rm F}$</annotation></semantics> </math> -FTP, <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>18</mn></msup> <mi>F</mi></mrow> <annotation>$^{18}{\rm F}$</annotation></semantics> </math> -FDG) to assess the impact on Alzheimer's disease (AD) classification. All harmonization procedures were conducted using leave-one-out cross-validation or by training on unpaired data and applying the learned transformations to paired data. PETHarmony effectively eliminated voxel-level discrepancies between PET/CT and PET/MR images ( <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics><mrow><mi>P</mi> <mo><</mo> <mn>0.05</mn></mrow> <annotation>$P < 0.05$</annotation></semantics> </math> reduced to n.s.). Phantom validation demonstrated that recovery coefficient curves were restored and closely aligned with the reference line, indicating improved quantitative accuracy. For cross-tracer consistency, linear regression between <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>18</mn></msup> <mi>F</mi></mrow> <annotation>$^{18}{\rm F}$</annotation></semantics> </math> -FBP and <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>11</mn></msup> <mi>C</mi></mrow> <annotation>$^{11}{\rm C}$</annotation></semantics> </math> -PiB was markedly improved toward the line of identity (y = x, R <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics><msup><mrow></mrow> <mn>2</mn></msup> <annotation>$^{2}$</annotation></semantics> </math> = 1). Specifically, in the GAAIN cohort, the regression line improved from y = 0.52x + 0.52, <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics><msup><mi>R</mi> <mn>2</mn></msup> <annotation>${\rm R}^{2}$</annotation></semantics> </math> = 0.89 to y = 0.93x + 0.13, R <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics><msup><mrow></mrow> <mn>2</mn></msup> <annotation>$^{2}$</annotation></semantics> </math> = 0.97. In the OASIS cohort, it improved from y = 0.51x + 0.55, <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics><msup><mi>R</mi> <mn>2</mn></msup> <annotation>${\rm R}^{2}$</annotation></semantics> </math> = 0.87 to y = 0.95x + 0.06, <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics><msup><mi>R</mi> <mn>2</mn></msup> <annotation>${\rm R}^{2}$</annotation></semantics> </math> = 0.95. Furthermore, PETHarmony improved multicenter AD classification accuracy by 15.3% ( <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>18</mn></msup> <mi>F</mi></mrow> <annotation>$^{18}{\rm F}$</annotation></semantics> </math> -FBP), 18.3% ( <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>18</mn></msup> <mi>F</mi></mrow> <annotation>$^{18}{\rm F}$</annotation></semantics> </math> -FTP), and 21.7% ( <math xmlns="http://www.w3.org/1998/Math/MathML"> <semantics> <mrow><msup><mrow></mrow> <mn>18</mn></msup> <mi>F</mi></mrow> <annotation>$^{18}{\rm F}$</annotation></semantics> </math> -FDG). PETHarmony achieves robust voxel-level harmonization of multicenter PET data, significantly improving cross-scanner and cross-tracer consistency and enhancing diagnostic accuracy. It provides a practical solution for standardizing quantitative PET in multicenter oncology, neuroscience, and other clinical trials.