A new algorithm that combines deep neural networks with trust region optimization can estimate the kinetic parameters of hepatocellular carcinoma from dynamic 18F-FDG PET/CT scans more accurately than established methods, according to a study published in EJNMMI Physics. The work, led by Jingchun Huang and colleagues, marks the first preliminary introduction of trust region Bayesian optimization, or TURBO, into pharmacokinetic analysis of PET data, and proposes a deep neural network-based variant the authors call DTRO.
The challenge is that estimating these parameters robustly is computationally difficult, and the authors note that while trust region optimization enables robust estimation in other domains, whether it could function in pharmacokinetic analysis remained unclear.
To test the approach, the researchers analyzed 28 hepatocellular carcinoma lesions in 24 patients, each of whom underwent a five-minute dynamic PET/CT scan followed by a one-minute static scan. The team compared parameters estimated by DTRO, TURBO and the conventional nonlinear least-squares method, using a two-input, three-compartment model as the reference framework. Performance was judged by how well each method’s parameters differentiated tumor tissue from background liver, and by root mean square error, R-squared values and receiver operating characteristic analysis after incremental learning.
DTRO achieved the lowest mean root mean square errors of the three methods, with values of 1.147 for hepatocellular carcinoma lesions and 1.009 for background liver tissue. Its mean R-squared values were also the highest, at 0.930 for tumors and 0.912 for liver. By comparison, nonlinear least squares produced mean RMSEs of 1.245 and 1.042 with R-squared values of 0.923 and 0.906, while TURBO yielded mean RMSEs of 1.200 and 1.025 with R-squared values of 0.928 and 0.909.
The diagnostic separation between tumor and healthy liver was also stronger with the new algorithm. DTRO significantly distinguished hepatocellular carcinoma from background liver with all six kinetic parameters, whereas TURBO achieved significant separation with four parameters, K1, k2, fa and vb, and nonlinear least squares did so with four parameters, k2, k4, fa and vb, all with P values below 0.05.
Receiver operating characteristic analysis reinforced that advantage. For the parameters K1 and k3, DTRO produced area under the curve values of 0.926 and 0.930, respectively, outperforming both nonlinear least squares, with AUCs of 0.887 and 0.805, and TURBO, with AUCs of 0.897 and 0.821.
The authors conclude that trust region optimization has proven feasible for pharmacokinetic estimation tasks, with the proposed DTRO algorithm achieving superior performance and offering a reliable foundation for hepatocellular carcinoma analyses.
Several caveats apply to the findings. The study was conducted on a small cohort of 24 patients with 28 lesions, and the article was shared early as a peer-reviewed, accepted manuscript subject to further edits before the final version of record. The results demonstrate estimation accuracy and diagnostic separation within this specific cohort and scanner protocol; broader clinical utility would require validation in larger, independent populations. The authors also describe the introduction of TURBO into this setting as preliminary, positioning DTRO as a foundation for future work rather than an immediately deployable clinical tool.
Related open-access articles in the same journal have explored multi-population multi-objective optimization for estimating 18F-FDG PET/CT pharmacokinetics of hepatocellular carcinoma, short-term kinetic estimation combining the maximum-slope method with a dual-input three-compartment model, and the relationship between PET findings and the Ki-67 proliferation index in tumor differentiation.
Subject of Research: A deep neural network-based trust region optimization algorithm for estimating 18F-FDG PET/CT kinetic parameters of hepatocellular carcinoma.
Article Title: Deep neural network-based trust region optimization for estimating the kinetic parameters of hepatocellular carcinoma from dynamic 18F-FDG PET/CT scans
Article References: Huang, J., Xiong, X., Lu, R., He, J., & Wang, S. (2026). Deep neural network-based trust region optimization for estimating the kinetic parameters of hepatocellular carcinoma from dynamic 18F-FDG PET/CT scans. EJNMMI Physics. https://doi.org/10.1186/s40658-026-00942-9
Image Credits: AI Generated
DOI: 10.1186/s40658-026-00942-9
Keywords: hepatocellular carcinoma, dynamic PET/CT, 18F-FDG, pharmacokinetic modeling, deep neural networks, trust region optimization, TURBO, compartment model, ROC analysis, quantitative imaging, EJNMMI Physics, Deep
News Source: Nathaniel Bowman. (October 29, 2026). Deep Learning Boosts PET Kinetic Parameter Estimation for Liver Cancer. Scienmag.
