AI tool maps patient-level dose rate from injected prostate cancer therapy

Clinical takeaway: Rapid, patient-specific maps could help specialists balance radiation delivered by the injected drug to tumors against exposure in the kidneys, liver, and other healthy tissue when planning later treatment cycles.

Imaging after radiopharmaceutical injection can show where the drug accumulated, but it does not by itself provide a patient-level map of how rapidly radiation is being absorbed across tumors and healthy organs. Producing that dose-rate map quickly enough to review before the next treatment cycle is the practical hurdle personalized dosing must clear.

Lutetium-177 prostate-specific membrane antigen (PSMA) therapy is an injected radiopharmaceutical treatment for metastatic castration-resistant prostate cancer. Unlike external-beam radiation, the radioactive drug circulates through the body and delivers radiation to PSMA-positive cancer sites. The amount of radioactivity injected and the treatment schedule are generally fixed, even though the radiation delivered to tumors and healthy organs varies within each patient and from one patient to another.

Monte Carlo radiation-transport simulations model how emitted particles move and deposit energy through patient-specific anatomy. They are the gold-standard reference for dose calculation but can take hours, while faster analytic methods sacrifice the spatial precision needed to assess organs at risk.

Investigators trained and evaluated an artificial intelligence model called DiffuDose in a secondary analysis of first-cycle single-photon emission computed tomography/computed tomography scans from 25 men receiving injected lutetium-177 PSMA therapy, using 18 patients for training and seven for validation. The model generated single-time-point dose-rate maps from scans obtained 24 to 216 hours after injection.

In 22.5 seconds, DiffuDose generated a full-resolution map estimating the radiation dose rate throughout the scanned anatomy after injection of the drug. Compared with Monte Carlo reference maps, it ranked first or tied for first on three of five image-level accuracy measures and second on the other two.

The six competing methods included an analytic Medical Internal Radiation Dose calculation and five artificial intelligence approaches: an encoder-decoder network, U-Net, a SwinIR transformer, a conventional diffusion model, and a latent diffusion model. DiffuDose’s estimates of average dose rate most closely matched the Monte Carlo calculations for both kidneys and the liver. The conventional and latent diffusion comparators required 400.3 and 160.2 seconds, respectively, although several non-diffusion methods were faster.

DiffuDose first generated a coarse dose-rate estimate from paired scans obtained after the radiopharmaceutical injection, then refined it with anatomical and functional detail to produce the final map. Investigators trained the model against patient-specific Monte Carlo calculations and assessed agreement across image slices and within segmented organs.

The team plans to use imaging at three time points after injection to create cumulative-dose reference maps, then train DiffuDose to estimate cumulative dose from a single post-injection scan. The authors also plan larger datasets, independent testing, and a fully three-dimensional version of the model.

“Right now, everybody gets the same dose,” Joyita Dutta, professor in the Riccio College of Engineering at UMass Amherst, said. “That essentially leaves the therapy’s potential untapped, to the extent that it’s suboptimal for a given patient. Measuring how much radiation each tissue actually absorbs is the key to personalizing treatment.”

Source: Lei B, et al. (2026 Jun 22) IEEE Trans Radiat Plasma Med Sci. DiffuDose: A Diffusion-Guided Model for Personalized Dosimetry for 177Lu-PSMA Radiopharmaceutical Therapy for Prostate Cancer