2036 - Toward Accurate PET SUV Assessment of Lung Cancer: Integrating Phantom-Calibrated Convolution Modeling with 4DCT Motion Compensation

Purpose/Objective(s): Positron emission tomography (PET) plays a crucial role in assessing tumor metabolic activity. However, its quantitative accuracy is often compromised by intrinsic uncertainties in signal detection and blurring caused by tumor motion. This study introduces a hybrid approach that combines phantom-calibrated Gaussian convolution modeling with 4DCT motion compensation to enhance standardized uptake value (SUV) quantification in lung cancer.

Materials/Methods: A capillary tube and a vial filled with ¹8F-FDG was imaged using a technology company scanner. The tube (50 mm in length, 0.5 mm inner diameter) served as a line source, and its cross-sectional profile was fitted to a Gaussian function to extract parameters for a convolution kernel, simulating detection uncertainty. The derived model was validated with the vial phantom before being applied to patient data. Ten lung cancer patients with well-defined tumors on CT were selected. PET/CT images were acquired using the same scanner in the phantom study, while 4DCT images were obtained during CT simulation. Gross tumor volumes (GTVs) were delineated in each phase to form the internal target volume (ITV). The 4DCT images were registered to PET/CT by aligning the tumor in PET/CT to the ITV. SUVs were generated for the GTVs based on the three models: (1) a uniform SUV distribution within the GTV, (2) a Gaussian kernel applied to each point in the GTV using sigma obtained from the phantom measurement, and (3) a single global Gaussian function across the entire GTV at its center. SUVs were summed over all GTVs to compute integrated SUVs in ITV and compared to actual PET-derived SUVs. Motion interpolation was performed when necessary. Profile analysis and histogram distributions were used for comparison.

Results: The sigma value obtained from the phantom study was 3.2 mm, consistent with PET resolution. Tumor motion ranged from 1.4 to 19.0 mm, with an average of 7.1 mm. The uniform SUV model (Model 1) showed the greatest deviation from the PET SUV, while the localized Gaussian convolution model (Model 2) improved agreement but tended to underestimate uptake at the center of the GTV. The best results were achieved with the global Gaussian spread model (Model 3). However, the amplitude and sigma of the Gaussian function must be optimized for each patient’s SUV data, as no universal sigma value could be determined across all patients in Model 3.

Conclusion: This study demonstrates that the calibrated Gaussian convolution model enhances SUV accuracy. However, it still exhibits relatively lower central uptake, which can be addressed by applying a patient-specific global Gaussian spread across the GTV. These findings suggest that a non-uniform model more accurately reflects the true tumor activity. Further refinement of the model is needed for better tumor characterization and potential improvement in personalized treatment planning and response assessment.