Abstract

Aim: This study aimed to develop an accurate and clinically useful framework for predicting organ-specific uptake, absorbed radiation dose, and treatment response of copper-64 (⁶⁴Cu), a theranostic radioisotope used for both cancer imaging and targeted therapy. Reliable prediction remains challenging due to patient variability and limited pharmacokinetic data.
Method: A hybrid computational framework was designed by integrating Monte Carlo radiation transport simulation (GATE v9.2), two-compartment pharmacokinetic modeling, and a deep learning CNN-LSTM network. The model was evaluated using clinical data from 47 patients diagnosed with glioblastoma multiforme (GBM), colorectal cancer, and non-small cell lung cancer (NSCLC). The dataset consisted of 1,240 in vitro uptake measurements, PET/CT time-activity curves acquired at seven time points, and 10 pharmacokinetic parameters. Performance was compared with Random Forest, XGBoost, and conventional LSTM models. SHAP analysis was applied for model interpretability.
Results: The proposed CNN-LSTM framework demonstrated strong predictive performance across all tasks. It achieved an AUC-ROC of 0.942 for treatment response classification, indicating high discrimination accuracy. For tumour absorbed dose estimation, the model produced an R² value of 0.918 with a mean absolute error of 1.21 Gy. Biodistribution prediction was also highly accurate, with an average error of 3.2%ID/g. These outcomes outperformed all benchmark machine learning models tested. SHAP interpretation identified SUVmax, tumour-to-background ratio, and the transfer rate constant k₁ as the most influential predictors.
Conclusion: The developed hybrid framework provides a robust, physics-consistent, and clinically useful tool for personalised ⁶⁴Cu dosimetry, treatment planning, and response prediction, pending prospective validation, with potential to improve precision oncology outcomes.

Key words: Absorbed dose prediction, biodistribution modelling, CNN-LSTM, ⁶⁴Cu-ATSM, dosimetry, GATE simulation, Monte Carlo