PET-CT based Dose Verification in Carbon Ion Radiotherapy for Choroidal Melanoma: A Linear Mixed-effects Model

Purpose: To quantify positron emission tomography-computed tomography (PET-CT) data analysis for verifying dose delivery in carbon-ion radiotherapy (CIRT) for choroidal melanoma (CM) by linear mixed-effects models (LMM) with machine learning. Methods: We aligned 3-dimensions (3D) dose and PET activity distributions in PET-CT images of 56 groups from 34 CM patients, establishing a direct correlation between planned doses and measured PET activity for reliable in-vivo dose verification. The dose conformity index (DCI) assessed the match between prescribed doses and the clinical target volume of the primary tumor (CTV _P ), while the PET-CT conformity index (PCI) measured how well the observed activity aligns with the target volume. The dose delivery index (DDI), defined as the PCI to DCI ratio, quantified this alignment. We applied a machine learning approach and a LMM to examine the correlation between irradiation dose and target activity in PET-CT images post-treatment. The Akaike information criterion (AIC) helped validate model accuracy, while the coefficient of determination (R²) determined dataset fit. To assess the generalizability of the proposed LMM, the external validation was performed using an independent cohort of 10 CM patients. Results: For isodose levels, the DDI values (DDI _95% =0.99±0.09, DDI _90% =0.98±0.13) were near 1, indicating high consistency between prescribed doses and observed PET activity. Each group’s PET 3D activity map included 19 isodose lines, and mean/median values of each map were analyzed with a LMM. The LMM showed strong predictive capability, with the mean model: A _Mean =(2.42+b _1i )*D _ij +(78.76+b _2i )*WTE _ij +(-55.90+b _0i )+e _ij where: b _0i ~N(0,4.22), b _1i ~N(0.0.80), b _2i ~N(0,92.81), and e _ij ~N(0,161.93) showed a significant fit with R² of 0.604, while the median model A _Median =(2.58+b _1i )*D _ij +(51.02+b _2i )*WTE _ij +(-56.46+b _0i )+e _ij where: b _0i ~N(0,65.71), b _1i ~N(0,0.93), b _2i ~N(0,0), and e _ij ~N(0.252.17) yielded a higher R² of 0.809. In the external validation, the prediction ability of PET values for mean-based model (R ² =0.96) was better than its for median-based model (R ² =0.90). Conclusion: PET-CT imaging in patients receiving CIRT for CM showed a measurable relationship between dose delivery and in-vivo imaging, confirming dose delivery accuracy and demonstrating the effectiveness of CIRT.