Discriminating inflammation from malignancy with short-dynamic patlak parametric 18 F-FDG PET/CT

Aim/Introduction Differentiating malignant from inflammatory uptake on 18 F-FDG PET/CT remains a major diagnostic challenge, as standardized uptake value (SUV) lacks specificity. Dynamic acquisitions with Patlak analysis can separate metabolized from unmetabolized tracer, potentially improving discrimination. We evaluated whether short-duration dynamic FDG PET/CT with Patlak parametric imaging provides complementary information beyond SUV for distinguishing malignancy from inflammation. Materials and Methods Twenty-seven patients undergoing oncologic PET/CT (breast, lung, or gastrointestinal cancer) were included, yielding 96 lesions (69 malignant, 27 inflammatory). Short dynamic acquisitions (20 min) were motion-corrected and analysed to generate influx rate ( K i ) and distribution volume ( V d ) maps. Lesions were segmented on SUV images (40% SUVmax), and radiomic features were extracted from SUV, K i , and V d maps. Exploratory data analysis, linear modelling, and dimensionality reduction assessed separability. A Random Forest classifier was trained with crossvalidation, integrating Synthetic Minority Oversampling (SMOTE) to address class imbalance. An independent validation cohort of 15 lesions (13 inflammatory, 2 malignant) was tested. Results Malignant lesions showed higher SUVmean (5.8 vs. 2.8 g/ml) and K i (1.95 vs. 0.75 ml/min/100ml), whereas inflammatory lesions demonstrated higher V d (44.7 vs. 35.1%). No single feature provided reliable thresholds. Logistic regression achieved 89% accuracy but suffered from quasi-separation, confirming limited linear discriminability. Random Forest classification yielded robust performance (cross-validated AUC-ROC 0.876; AUC-PR 0.948). With G-mean thresholding, inflammation was detected with high recall (0.93) but recall for malignancy was lower (0.74). Feature importance highlighted SUV and K i variance, as well as K i / V d ratios, as strongest predictors. In the external validation set, accuracy reached 0.80, with inflammation reliably identified (precision 0.85, recall 0.85). Conclusion Short dynamic Patlak imaging combined with machine learning improves the characterization of malignant versus inflammatory uptake beyond SUV alone. By decomposing FDG up-take into metabolized ( K i ) and unmetabolized ( V d ) fractions, this approach provides physiologically meaningful separation of tracer behaviour. While sensitivity for malignancy requires further optimization, our findings establish a reproducible framework for future more extensive research on clinical interpretation of parametric imaging in oncologic PET.