Abstract / Summary
Abstract Background : Multi-time-point kidney segmentation from Lutetium-177 (Lu-177) SPECT/CT is essential for patient-specific dosimetry for radiopharmaceutical therapy (RPT). However, expert segmentation remains time-consuming and subject to inter-observer variability, limiting its routine clinical use. Methods : This study compared the performance of four 3D deep-learning models including U-Net, Attention U-Net, UNETR, and SwinUNETR, for automated kidney segmentation and assessed their impact on internal dosimetry. A retrospective study was conducted using SPECT/CT datasets from patients who received Lu-177 PSMA or Lu-177 DOTATATE therapy. Performance was assessed using Dice Similarity Coefficient (DSC) and 95th percentile Hausdorff Distance (HD95). The best-performing model was applied to internal dosimetry calculations using MIRDcalc software. Results : SwinUNETR achieved the highest segmentation performance with DSC of 0.91±0.07 and HD95 of 12.50±14.82 mm. Dosimetry based on automated segmentation showed a mean absorbed dose difference of 10.28% relative to expert-based dosimetry. The DOTATATE group received mean kidney dose of 4.74±1.70 Gy/cycle, while the PSMA group received 1.73±1.89 Gy/cycle. Automated segmentation reduced processing time by 81-83% while maintaining accuracy comparable to expert segmentation. Conclusion : 3D deep-learning model, particularly SwinUNETR, demonstrates clinical feasibility for automated segmentation in Lu-177 SPECT/CT dosimetry workflows. Moreover, it significantly reduces clinician workload while maintaining clinically acceptable dosimetric accuracy.