FDG-PET/CT: Metabolic and Anatomical Imaging

• FDG-PET/CT is a hybrid imaging modality that fuses FDG-based metabolic activity with CT-derived anatomical details for precise oncologic evaluations.
• It employs automated segmentation and deep learning methods to extract quantitative biomarkers like MTV and SUV for robust tumor characterization.
• Parametric modeling and radiomics extraction from FDG-PET/CT enhance clinical decision-making by linking functional imaging with treatment outcomes.

Fluorodeoxyglucose positron emission tomography / computed tomography (FDG-PET/CT) is an integrative imaging modality that combines functional detection of glucose metabolism via radioactive 18F-fluorodeoxyglucose (FDG) with cross-sectional anatomical reconstruction from computed tomography (CT). FDG accumulates in malignant tissues due to upregulated glycolysis, and CT facilitates anatomical localization, attenuation correction, and discrimination between pathological and physiological uptake. Automated lesion segmentation, parametric analysis, and radiomic quantification are central to FDG-PET/CT’s application in oncologic decision-making, protocol development, and biomarker research. Robust methods for extracting metabolic tumor volume (MTV), kinetic parameters, and derived imaging features have become widely adopted in both academic and clinical workflows (Sibille et al., 2022, Xue et al., 5 Nov 2025, Scussolini et al., 2017).

1. Principles of FDG Metabolism and PET/CT Imaging

FDG is a glucose analog radiolabeled with 18F. After injection, FDG is taken up by cells via GLUT transporters and phosphorylated by hexokinase to FDG-6-phosphate (Cai et al., 2023). Unlike glucose-6-phosphate, FDG-6-phosphate is metabolically trapped, yielding a time-integrated measure of glycolytic activity—a hallmark of neoplastic transformation. PET detects the 511 keV annihilation photons emitted during positron decay, reconstructing 3D maps of FDG distribution. CT simultaneously produces high-resolution anatomical images and tissue attenuation maps at ~70 keV, which are converted for attenuation correction of PET data.

Quantification is standardized via the Standardized Uptake Value (SUV):

$$\mathrm{SUV} = \frac{C_{\rm tissue}(t)\,\text{[kBq/mL]}}{A_{\rm inj}\,\text{[kBq]}/W\,\text{[g]}}$$

where $C_{\rm tissue}(t)$ is the activity concentration at time $t$, $A_{\rm inj}$ is the decay-corrected injected dose, and $W$ is the patient’s body weight (Xue et al., 5 Nov 2025). SUV metrics include SUVmax (maximum per lesion), SUVmean, and SUVpeak. Attenuation correction relies on the voxel-wise factor:

$$\mathrm{ACF}(x,y,z) = \exp\left(\int_{\text{path}} \mu(s)\,\mathrm{d}s\right)$$

where $\mu(s)$ is the CT-derived linear attenuation coefficient (Cai et al., 2023).

2. Segmentation, Biomarker Extraction, and Quantitative Performance

Automated segmentation is critical for extracting quantitative biomarkers such as MTV—the total lesion volume above an uptake threshold. MTV is a robust surrogate for tumor burden and is predictive of prognosis, treatment response, and progression-free survival (Sibille et al., 2022). Manual segmentation suffers from high inter-reader variability and is labor-intensive; deep learning methods (U-Nets, nnUNet, transformers) now dominate research and challenge datasets.

A representative approach employs a cascaded ensemble of 3D UNet models at downsampled 6 mm resolution, processing five channels (SUV-normalized PET, CT, CT_Soft, CT_Lung, SUV_hot), followed by a residual refiner network to upsample and refine masks at the original resolution. Ensemble outputs from four UNet variants (15-fold stratified splits) are linearly fused, and the final segmentation is post-processed (Sibille et al., 2022).

Segmentation accuracy is assessed by Dice Similarity Coefficient:

$$\mathrm{Dice} = \frac{2\,|S_{\rm auto} \cap S_{\rm gt}|}{|S_{\rm auto}| + |S_{\rm gt}|}$$

On a held-out test set, Dice = 0.68 and $R^2$ for MTV correlation between manual and automatic methods is 0.969, with slope = 0.947, reflecting near-linear agreement and limited systematic bias. Inference time averages 89.7 s per case on an NVIDIA V100 (Sibille et al., 2022).

3. Datasets, Protocols, and Preprocessing Standards

Large-scale, well-annotated datasets provide the foundation for benchmarking and model development. PETWB-REP, for example, includes 490 curated whole-body FDG-PET/CT scans and paired radiology reports spanning lung, liver, breast, prostate, and ovarian cancer. Acquisition protocols—3.7–5.55 MBq/kg FDG, 60-min uptake, 120 kVp/170 mA CT—are standardized, and PET is reconstructed via OSEM (2 iterations, 21 subsets, 5 mm Gaussian filter). Quality control comprises physician visual review and automated NIfTI/DICOM conversion integrity checks (Xue et al., 5 Nov 2025).

Preprocessing pipelines entail attenuation and scatter correction, SUV conversion per voxel, normalization (z-score for CT, linear scaling for PET), and co-registration to a shared in-plane resolution (0.98 mm) and slice thickness (3.0 mm). Data structures organize PET_RAW, CT_RAW, normalized volumes, and metadata (patient ID, cancer type, dose, uptake time, weight). These protocols facilitate radiomics extraction, multi-modal and NLP research, and robust cross-institutional transfer.

4. Mathematical Modeling and Parametric Analysis

FDG kinetic analysis mathematically models tracer exchange using compartmental ODEs—typically two- or three-compartment models for different tissue types (Scussolini et al., 2017, Piana et al., 2020). For instance, hepatic FDG fate is governed by:

$$\begin{cases} \frac{dC_f}{dt}(t) = k_{fa}\,C_a(t) - (k_{af} + k_{mf})\,C_f(t) + k_{fm}\,C_m(t) \ \frac{dC_m}{dt}(t) = k_{mf}\,C_f(t) - k_{fm}\,C_m(t) \end{cases}$$

where $C_a$ is arterial input, $C_f$ free FDG, $C_m$ metabolized pool, and $k_{ij}$ rate constants. Voxel-wise parametric maps are estimated by nonlinear least squares with regularization (e.g., Gauss–Newton optimization with Tikhonov penalty):

$$J(\mathbf{k}) = \sum_{i=1}^N \|C_{\mathrm{measured}}(t_i) - C_{\mathrm{model}}(t_i;\mathbf{k})\|^2 + \lambda R(\mathbf{k})$$

Parametric imaging reveals regional heterogeneity, with coefficient-of-variation reduced ~30–50% compared to ROI-level fitting (Scussolini et al., 2017).

In cardiac PET, quantification of myocardial blood flow (MBF) is feasible by extracting FDG delivery rate $K_1$ and normalizing for blood glucose levels. The generalized Renkin–Crone model:

$$K1_{FDG} = MBF \cdot [1 - a_{FDG} \exp(-b_{FDG}/MBF)]$$

yields MBF estimations (r > 0.9 with reference 82Rb flow after normalization) (Zuo et al., 2020).

5. Network Architectures and Multimodal Integration

Recent advances deploy UNet, ResUNet (ResNet-18/50 backbone), cascaded low/high-resolution models, and transformer-based networks (Hyper-Connected Transformer [HCT], SwinUNet3D) for PET/CT fusion and lesion segmentation (Bi et al., 2022, Guha et al., 6 Jan 2026). Multi-branch encoders process PET, CT, and concatenated channels, using self-attention to integrate local and global features. HCT, for example, combines modality-specific ResNet-50 patch embeddings, multi-head self-attention, and iterative hyper-connected fusion via transformer decoders. This approach improves Dice scores—HCT yields 72.3% (NSCLC) and 66.4% (sarcoma), outperforming vanilla CNNs (Bi et al., 2022). SwinUNet3D achieves Dice = 0.88 (IoU = 0.78) versus 3D U-Net Dice = 0.48, with superior detection of small and irregular lesions and reduced false positives (Guha et al., 6 Jan 2026).

Optimization strategies include combined losses (Dice, cross-entropy, and sensitivity terms), spatial weighting based on localization cues (as in L2SNet), and post-processing (small connected component removal). Adaptive thresholding, sliding-window fusion, and test-time augmentation further refine predictions and generalizability (Hadlich et al., 2023, Amiri et al., 2022).

6. Clinical Applications, Radiomics, and Future Directions

Automated FDG-PET/CT segmentation drives reproducible MTV quantification, volumetric burden assessment, and therapy response monitoring—reducing manual contouring time and inter-reader variability (Sibille et al., 2022). Radiomics pipelines extract hundreds to thousands of features (shape, histogram, GLCM, run-length, zone-size) from manually or automatically segmented lesions, with standardized resampling and binning (typically 1 mm³ isotropic voxels, bin = 5 SUV units) (Henriquez et al., 2023). These features enable non-invasive prediction of clinically relevant mutations, such as EGFR in NSCLC (AUC ≈ 0.82 for best models), and response/prognosis stratification (Henriquez et al., 2023, Xue et al., 5 Nov 2025).

Parametric imaging supports functional phenotyping (e.g., identifying regions of FDG washout due to inflammatory infiltrates by voxel-wise regression, relevant for immunotherapy monitoring) (Parodi et al., 2024). Advanced quantification in bone marrow regions now incorporates dual-energy CT–derived bone-fraction correction (BFC) to account for trabecular bone admixture and yield SUV, $K_1$, and $K_i$ increases (~13.3% on average), reflecting true marrow metabolism (Li et al., 29 Dec 2025).

Future work prioritizes multi-scale hybrid architectures, spatial and anatomical priors, uncertainty quantification, and model robustness across scanner types and acquisition protocols (Sibille et al., 2022, Guha et al., 6 Jan 2026). Expansion to multi-tracer PET, large multicenter cohorts, and integration with clinical PACS and report generation are ongoing directions (Xue et al., 5 Nov 2025).

7. Limitations and Considerations

Despite significant advances, automated FDG-PET/CT methods may miss sub-centimeter or low-avid lesions (false negatives), misclassify physiologically or inflamed regions as tumor (false positives), and be sensitive to scanner variability and reconstruction parameters (Sibille et al., 2022). Hybrid architectures and transformer networks partially address these concerns, yet challenges persist in generalization across domains and acquisition protocols (Alloula et al., 2023). The accuracy and interpretability of radiomics-derived predictions are contingent on robust segmentation, standardized preprocessing, and multi-center harmonization. Compartmental modeling and parametric analysis require high SNR dynamic data, accurate input functions, and careful handling of regularization and identifiability (Scussolini et al., 2017, Piana et al., 2020). Clinical adoption relies on transparent validation, reproducibility, and regulatory-compliant pipelines.

---

FDG-PET/CT represents a convergent platform integrating metabolic and anatomical information for oncology imaging, quantitative biomarker extraction, radiomics, and image-based therapy planning. Recent developments in deep segmentation, transformer modeling, parametric imaging, and robust protocol design have established FDG-PET/CT as a central methodology in cancer research and clinical care, with ongoing advances addressing remaining limitations in specificity, generalization, and interpretability (Sibille et al., 2022, Xue et al., 5 Nov 2025, Guha et al., 6 Jan 2026, Li et al., 29 Dec 2025).