Differential performance of statistical versus machine learning methods in partial volume correction in oncologic
Hanan K Metwally1, Hesham A Gawaad2, Magdy M Khalil3
1Medical Biophysics, Department of Physics, Faculty of Science, Helwan University, Cairo, Egypt.
Annals of Nuclear Medicine
|July 31, 2026
Summary
Machine learning significantly improves partial volume correction (PVC) in positron emission tomography (PET) imaging, outperforming traditional methods. This enhances PET quantification accuracy for better lesion characterization and therapy assessment.
Area of Science:
- Medical Imaging
- Nuclear Medicine
- Computational Biology
Background:
- Partial volume effects (PVE) in PET imaging cause quantification inaccuracies, especially in small or heterogeneous lesions.
- Accurate quantification is crucial for lesion characterization, therapy response assessment, and multi-center studies.
Purpose of the Study:
- To systematically evaluate statistical and machine learning (ML) approaches for partial volume correction (PVC) in PET.
- To compare PVC performance across different PET reconstruction algorithms (TrueX, TrueX+TOF, Iterative+TOF).
Main Methods:
- Phantom experiments were conducted with varying lesion sizes and contrast ratios.
- Exponential recovery coefficient (RC) fitting models were derived and validated.
- Machine learning algorithms (Random Forest, Support Vector Regression, Gradient Boosting) were implemented for PVC.
Main Results:
- Iterative+TOF reconstruction provided the most consistent RC estimates.
- ML-based PVC significantly reduced residual errors compared to exponential fitting.
- Random Forest demonstrated the best performance with the lowest RMSE and highest R².
Conclusions:
- Machine learning-driven PVC offers superior accuracy and robustness for PET quantification.
- ML-based PVC has the potential to standardize PET data and improve clinical assessments.
- This approach can enhance lesion characterization and multi-center data harmonization in PET imaging.
