Voxel-wise adjustment of single-time-point dosimetry using deep learning for 177Lu-PSMA radiopharmaceutical therapy

Introduction Single-time-point (STP) dosimetry is practical for clinical routine but remains limited by population-based kinetic assumptions and strict late imaging time points. This study introduced a deep learning (DL) method for voxel-wise adjustment of STP dosimetry in 177 Lu-PSMA radiopharmaceutical therapy (RPT), aiming to improve accuracy and scan-time flexibility. Methods A total of 317 post-therapeutic SPECT/CT scans from 21 patients with metastatic castration-resistant prostate cancer (mCRPC) receiving multiple cycles of [ 177 Lu]Lu-PSMA-617 RPT were included. A decomposition-based DL model was developed to predict a voxel-wise scaling factor map from single-time-point SPECT/CT. The predicted map was multiplied with the quantified SPECT and a learned polynomial of imaging time to generate the corrected TIA map. Ground-truth TIA maps were generated using voxel-wise integrals of time-activity curves fitted by sums of exponential functions. Patient-wise five-fold cross-validation and comparative experiments were performed. Generalization was evaluated using an external dataset of 23 patients with mCRPC receiving [ 177 Lu]Lu-PSMA-I&T RPT. Results Across all test data (n=317), the average voxel-wise normalized root-mean-square errors were within 15.0% for all evaluated organs and tumors, with a global structural similarity of 0.9691. Using 24h SPECT/CT (n=81), the method achieved mean absolute percentage errors of 17.9%±16.5% for kidneys, 15.8%±13.4% for liver, 21.0%±28.8% for spleen, 20.6%±18.2% for bone marrow, and 26.5%±19.1% for tumors. The MAPEs generally decreased with later imaging, reaching 10.3%±9.1% (kidneys), 10.2%±7.8% (liver), 21.3%±12.9% (spleen), 11.1%±8.7% (bone marrow), and 13.3%±11.2% (tumors) at 72h (n=49). On the external dataset, the lowest MAPEs were 18.6%±12.0% for liver and 23.7%±20.9% for spleen at 24h, and 12.6%±14.4% for kidneys and 20.2%±12.2% for bone marrow at 48h. Conclusion The decomposition-based DL method enabled voxel-wise correction of STP dosimetry at time points including 24h, 48h, and 72h post injection, improved scan-time flexibility, and demonstrated the technical feasibility of instant STP dosimetry.