TL;DR
This paper introduces an unpaired, physics-informed CycleGAN that simultaneously estimates the arterial input function (AIF) and voxel-wise kinetic parameter maps directly from dynamic PET scans. By integrating the analytical two-tissue compartment model into the cycle-consistency loop, the network avoids both invasive arterial blood sampling and computationally prohibitive non-linear fitting algorithms.
What Problem Does This Paper Address?
Dynamic positron emission tomography (dPET) provides invaluable functional insights into metabolic processes—such as glucose metabolism in tumors—that static PET cannot reveal. Extracting these kinetic parameters requires fitting a tracer kinetic model to time-activity curves (TACs) across tissues.
However, standard kinetic modeling suffers from two major roadblocks:
- The AIF bottleneck: Fitting requires knowing the tracer concentration in arterial blood over time. The gold standard is continuous arterial blood sampling during the scan, which is invasive, labor-intensive, and rarely practical in routine clinical settings.
- Computational overhead: Solving non-linear compartment models voxel-by-voxel across a 3D volume is slow and sensitive to image noise. Consequently, whole-body dynamic scans are rarely analyzed at the voxel level.
Methods
The authors adapt a physics-informed CycleGAN framework (originally proposed for dynamic contrast-enhanced MRI) to dPET imaging:
- Unpaired Mapping: The model does not require perfectly paired dPET scans and parametric ground truths during training. Unpaired training samples a dPET scan from one subject and a kinetic parameter map (with corresponding AIF) from another.
- Generator ($G$): A feedforward neural network takes dynamic 3D PET time frames (42 input channels) and directly outputs 4 kinetic parameter maps ($K_1, k_2, V_b, k_3$) alongside the estimated AIF.
- Physics-Based Forward Model ($F$): Instead of using a learned neural network to translate kinetic maps back into dPET images, the reverse path uses an exact, analytical irreversible two-tissue compartment model (2TCM). The model explicitly convolves the estimated AIF with tissue impulse responses using the predicted rate constants.
- Cycle Consistency & Adversarial Loss: The generator must produce parameters that accurately reconstruct the observed PET time-series through the known physical laws, while a PatchGAN discriminator ensures the generated parameter maps match realistic distributions.
dPET Image (X) ──[ Generator G ]──> Parameters (Y) + AIF (C_A)
▲ │
└──────[ Analytical 2TCM (F) ]────────────┘
Datasets
- Preclinical Mouse Cohort: Evaluated on 70 whole-body $[^{18}\text{F}]\text{FDG}$ dynamic PET scans of mice.
- Acquisition Details: Scans included simultaneous invasive arterial blood sampling as the reference standard.
- Image Dimensions: Dynamic volumes have dimensions $42 \times 96 \times 48 \times 48$ (time frames $\times$ axial $\times$ coronal $\times$ sagittal). Time framing consisted of 42 frames: $1 \times 30\,\text{s}$, $24 \times 5\,\text{s}$, $9 \times 20\,\text{s}$, and $8 \times 300\,\text{s}$.
- Reference voxel-wise kinetic maps were generated using a linearized 2TCM fit against the true blood samples.
Key Insights and Ideas
- Physics replaces neural inversion: Replacing the backward generator with the deterministic 2TCM equation eliminates half of the typical CycleGAN instability. The laws of tracer kinetics act as an explicit regularizer.
- Unified feedforward inference: Instead of first predicting the AIF and then numerically fitting kinetic curves across millions of voxels, inference collapses into a single forward pass through the generator network.
- Competitive with ensembled baselines: Even without supervised pairs, a single CycleGAN model rivals dedicated, supervised deep learning models that rely on 10-model ensembling.
Key Results
- Parameter Map Fidelity: On a held-out test set, the generated kinetic parameter maps achieved a Peak Signal-to-Noise Ratio (PSNR) of 34.49 dB and a Structural Similarity Index (SSIM) of 0.7425 relative to reference linearized fits.
- AIF Accuracy: The CycleGAN's estimated input function achieved a linear regression fit of $y = 0.91x + 0.21$ with a Pearson correlation of $r = 0.90$ against sampled blood curves.
- Comparison to Supervised Ensembles: Performance came close to dedicated, supervised deep learning baselines:
- FC-DLIF (10-model ensemble): $y = 0.93x + 0.12$, $r = 0.96$
- PI-DLIF (10-model ensemble): $y = 0.91x + 0.09$, $r = 0.95$
Limitations and Future Work
- Generator Normalization Artifacts: The current architecture uses instance normalization, which normalizes features per-volume. The authors note this interferes with recovering the absolute scale of physiological rate constants.
- Limited Preclinical Scope: The model was trained and tested solely on 70 healthy mice; validation on human clinical data and pathology (e.g., heterogeneous tumors) is still needed.
- Lack of Uncertainty Estimation: Because the network performs generative reconstruction, incorporating explainable AI (XAI) and voxel-level uncertainty estimation will be vital to catch hallucinated kinetics or out-of-distribution failures before clinical deployment.
Summary updated