Radiopharmaceutical therapy (RPT) is entering a new era in cancer treatment, but one of its biggest limitations has remained stubbornly unchanged: patients generally receive standardized doses rather than amounts tailored to how their individual bodies absorb radiation. Researchers at the University of Massachusetts Amherst have now developed an artificial intelligence system that could help change that. Called DiffuDose, the model generates detailed, patient-specific radiation dose maps in less than 23 seconds, achieving accuracy comparable to a computational technique considered the current gold standard.
The advance could be especially important for men receiving radiopharmaceutical therapy for advanced prostate cancer. The treatment received U.S. Food and Drug Administration approval in 2022 and uses a radioactive drug linked to a molecule that seeks out prostate-specific membrane antigen, or PSMA, found on many prostate cancer cells. The resulting therapy, based on the radioisotope lutetium-177, or 177Lu, travels through the bloodstream and deposits radiation in tumors throughout the body. Unlike external-beam radiation, which targets a defined region from outside the body, RPT can reach cancer that has spread to multiple locations.
Its systemic reach is also the source of one of its central dangers. The radioactive compound can accumulate in healthy tissues, including the kidneys and liver. The same radiation that damages cancer cells can injure normal organs when exposure becomes too high. “The main issue with many cancer treatments is toxicity,” says Joyita Dutta, a professor in the Riccio College of Engineering at UMass Amherst and senior researcher on the project. “Whether it’s radiation, chemo, or radiopharmaceutical therapy—whatever mechanism kills the cancer cells also hurts healthy tissue.”
Doctors can estimate this risk through dosimetry, the process of determining how much radiation has been absorbed by specific tissues. After treatment, medical scans show where the radiopharmaceutical has concentrated, but images alone do not directly reveal the absorbed dose. Dosimetry requires additional computational analysis that combines information about the radioactive material’s distribution, its physical decay and the way emitted particles travel through tissue. The most accurate existing calculations can take hours for a single patient, limiting their usefulness in routine clinical decision-making.
DiffuDose was designed to make that analysis fast enough to support treatment planning. The system combines two artificial intelligence modules. The first produces a coarse estimate of the radiation distribution, rapidly identifying the broad pattern of dose across the body. A second module then refines that estimate, reconstructing a full-resolution map that captures finer variations from one region to the next. The approach uses diffusion-guided deep learning, a class of generative modeling techniques that can progressively refine an initial prediction toward a more detailed and realistic result.
The output is not simply a single number describing a patient’s total radiation exposure. It is a three-dimensional, pixel-by-pixel representation of how the dose is distributed throughout the body. Such a map can show whether a tumor is receiving substantial radiation while nearby organs remain below a potentially dangerous threshold. That distinction is essential because two patients given the same administered activity may absorb very different amounts of radiation depending on anatomy, blood flow, tumor burden, kidney function and the way the radiopharmaceutical is cleared.
In testing, the UMass Amherst team compared DiffuDose with six other computational approaches. The new model delivered the strongest overall quantitative performance and maintained consistently high accuracy across several organs, including both kidneys and the liver—organs that are particularly important when assessing toxicity in 177Lu-PSMA therapy. Most notably, it generated results matching the accuracy of the gold-standard method in under 23 seconds per patient, a dramatic reduction from the hours required by conventional calculations.
Rapid dosimetry could eventually give oncologists more flexibility in deciding how much treatment a patient can safely receive and when another cycle should be administered. If a patient’s organs tolerate an initial treatment well, a subsequent dose might be adjusted upward; if radiation accumulates in a vulnerable organ, treatment could instead be reduced, delayed or reconsidered. The system is not intended to replace clinical judgment, and its performance will need to be validated in broader patient populations and prospective clinical studies before it can guide routine care. Its immediate significance is that it demonstrates how computational methods might remove a major practical barrier to individualized RPT.
Dutta’s group is now pursuing collaborations with UMass Chan Medical School to develop models that combine post-treatment imaging with biological information, including blood biomarkers. The goal is to determine not only where radiation is deposited, but also how a patient’s disease is responding. The project involved collaborators Michael King at UMass Chan Medical School, Alejandro Bertolet and Quanzheng Li at Massachusetts General Hospital, and Babak Saboury at the Institute of Nuclear Medicine in Bethesda, Maryland. The findings were reported in IEEE Transactions on Radiation and Plasma Medical Sciences under the title “DiffuDose: A Diffusion-Guided Model for Personalized Dosimetry for 177Lu-PSMA Radiopharmaceutical Therapy for Prostate Cancer.” If further validated, the technology could help transform RPT from a largely standardized treatment into a more responsive therapy calibrated to each patient’s unique radiation biology.
Subject of Research: People
Article Title: DiffuDose: A Diffusion-Guided Model for Personalized Dosimetry for 177Lu-PSMA Radiopharmaceutical Therapy for Prostate Cancer
News Publication Date: 22-Jun-2026
Web References: University of Massachusetts Amherst; IEEE Xplore: https://ieeexplore.ieee.org/document/11573134
References: IEEE Transactions on Radiation and Plasma Medical Sciences. DOI: 10.1109/TRPMS.2026.3705613
Keywords: Radiopharmaceutical therapy, prostate cancer, 177Lu-PSMA, personalized dosimetry, artificial intelligence, deep learning, diffusion models, radiation dose mapping, medical imaging, cancer treatment, precision medicine, radiation toxicity
