AI segmentation to make voxel-level radioligand dosimetry practical at scale

Personalised, voxel-level dosimetry after radioligand therapy requires segmenting organs and tumours on serial SPECT/CT or PET/CT scans, a step that traditionally takes a trained physicist or physician several hours per patient using manual contouring. Deep-learning segmentation models, typically convolutional neural networks such as U-Net or newer transformer architectures, have been developed to automate this step. A 2025 comparison of AI-based against fully manual kidney segmentation for single-time-point 177Lu-PSMA dosimetry, using 8 treatment cycles from 5 patients, found no statistically significant difference in mean absorbed dose between the two methods, with a Dice similarity coefficient of 0.898 between AI and manual contours, though the underlying kidney volumes did differ significantly between methods.

A broader review of automated dosimetry pipelines reports that combining CNN-based organ segmentation with automated registration and dose-rate estimation can cut total per-patient processing time from 4-6 hours to 30-45 minutes, a 70-90% reduction, while achieving Dice similarity coefficients of 0.88-0.95 against expert manual contours for organ segmentation specifically. Performance for tumour lesion segmentation, as opposed to organ segmentation, is less mature: an external, held-out-centre test of an nnU-Net lesion segmentation model for PSMA PET/CT and post-therapy SPECT/CT reported lower Dice scores of 0.76-0.78, and the model's PET-derived tumour volume estimates were described as unstable in that external test.

An independent 2026 evidence review of AI in oncologic theranostics concludes that, despite this technical progress, no AI segmentation or dose-prediction model for PSMA or peptide receptor radionuclide therapy dosimetry has yet been validated in a prospective, multicentre trial demonstrating that AI-guided activity adaptation (continuing, delaying or stopping treatment based on the AI-derived dose) actually changes patient outcomes; the review explicitly calls this the priority next study for the field.