A new AI model has generated patient-specific radiation dose-rate maps in about 23 seconds, potentially helping make personalised dosing more practical for radiopharmaceutical cancer therapy.

Radiopharmaceutical therapy (RPT) is becoming an increasingly important approach to treating cancer, using radioactive medicines to deliver radiation directly to tumours and other cancer targets throughout the body. However, determining how much radiation individual tissues actually receive remains a challenge.

Researchers at the University of Massachusetts Amherst have developed DiffuDose, an AI-driven framework designed to rapidly estimate voxel-level radiation dose from post-treatment SPECT and CT images.

In a study published in IEEE Transactions on Radiation and Plasma Medical Sciences, the model was evaluated using a clinical cohort receiving 177Lu-PSMA radiopharmaceutical therapy for metastatic prostate cancer.

The researchers found that DiffuDose produced dose-rate maps with accuracy comparable to Monte Carlo dosimetry, regarded as the gold-standard computational approach for dose calculation, while substantially reducing the computational time required.

The model achieved an inference time of approximately 22.5 seconds per patient, compared with more than 400 seconds for the Monte Carlo implementation used for comparison.

The findings could support the development of more practical personalised dosimetry approaches for radiopharmaceutical therapy, although further research will be needed before the technology can be used to guide treatment decisions in routine clinical practice.

AI targets a challenge in radiopharmaceutical therapy

Unlike conventional external-beam radiotherapy, which delivers radiation from a machine outside the body, RPT uses a radioactive drug that travels through the body and accumulates in targeted tissues.

177Lu-PSMA therapy is used to treat prostate cancer by targeting prostate-specific membrane antigen (PSMA), which is expressed on prostate cancer cells. The approach allows radiation to be delivered to distributed cancer sites, including metastatic lesions.

However, the radiation emitted by the treatment can also affect healthy organs. Understanding the amount and distribution of radiation absorbed by tumours and organs at risk could therefore be important for balancing therapeutic effects against toxicity.

This process is known as dosimetry.

Current approaches involve trade-offs between speed and accuracy. Monte Carlo simulations can provide highly accurate dose calculations by modelling radiation transport, but their computational demands can make them difficult to incorporate into routine clinical workflows.

Faster analytical approaches, including methods based on the Medical Internal Radiation Dose (MIRD) formalism, can provide rapid estimates but may lack the spatial precision required for detailed voxel-level dose assessment.

The researchers developed DiffuDose to address this gap.

The framework combines a diffusion probabilistic model with multiscale information from SPECT and CT imaging. Its architecture includes a low-resolution generator that produces an initial dose-rate estimate and a super-resolution refinement module that generates a full-resolution output.

The resulting maps provide a voxel-level view of how radiation is distributed throughout the body.

Joyita Dutta, professor at the University of Massachusetts Amherst and senior author of the study, said: “Measuring how much radiation each tissue actually absorbs is the key to personalizing treatment.”

DiffuDose matches Monte Carlo accuracy with lower computational demands

The researchers compared DiffuDose with established analytical calculations and several deep learning and diffusion-based approaches.

Across quantitative measures, DiffuDose performed strongly against the competing methods and produced results close to the Monte Carlo reference.

The model achieved a peak signal-to-noise ratio of 44.40 dB, compared with 36.12 dB for MIRD calculations, while its structural similarity score was 0.961. Its relative absolute error was 4.64%, compared with 5.70% for MIRD.

The researchers also assessed the model’s performance across organs including the kidneys and liver, which are among the tissues that can be at risk of radiation toxicity.

Visual comparisons showed that DiffuDose was able to reproduce both broader radiation distributions and smaller focal areas in the dose maps. The researchers also assessed dose-volume characteristics, finding that the predicted distributions were consistent with the Monte Carlo reference.

The speed of the model could be particularly relevant to the future use of personalised dosimetry.

If accurate dose maps can be generated rapidly, clinicians could potentially use information about how an individual patient’s tumour and healthy tissues respond to treatment when considering subsequent therapy.

However, the current study does not demonstrate that DiffuDose has been used prospectively to personalise treatment or improve clinical outcomes.

Instead, the work demonstrates the technical feasibility of using AI to generate accurate dose-rate maps rapidly from post-therapy imaging.

AI could help move RPT beyond one-size-fits-all dosing

The researchers see the technology as part of a broader move towards personalised radiopharmaceutical therapy.

At present, radiopharmaceutical treatments are generally administered according to standardised protocols rather than being individually adjusted according to the precise radiation dose absorbed by each patient’s tumours and organs.

Personalised dosimetry could potentially provide a more detailed picture of individual radiation exposure, helping researchers investigate whether treatment dose and frequency could eventually be adapted according to patient-specific characteristics.

The potential importance of this approach is growing as radiopharmaceutical therapy expands beyond a small number of established treatments.

The researchers’ next steps include using multiple post-treatment SPECT scans to estimate cumulative radiation dose, developing a fully three-dimensional version of the model and investigating whether accurate dosimetry could be performed using a single post-treatment scan.

They also plan to explore the application of DiffuDose to other radiopharmaceutical therapies, including 177Lu-DOTATATE and alpha-particle therapies.

These developments could broaden the potential application of AI-based dosimetry beyond 177Lu-PSMA therapy for metastatic prostate cancer.

Dutta’s research team is also exploring collaborations to combine post-treatment imaging with patient blood biomarkers, with the aim of better understanding individual responses to radiopharmaceutical therapy.

For now, DiffuDose remains a research-stage computational tool rather than a clinically validated system for directing patient treatment. The study nevertheless highlights how AI could help address one of the practical barriers to personalised radiopharmaceutical therapy: generating accurate radiation dose maps quickly enough to fit into future clinical workflows.

As radiopharmaceutical therapies continue to develop, technologies capable of rapidly quantifying where radiation is delivered — and how much is absorbed by individual organs and tumours — could become increasingly important in efforts to balance treatment efficacy with toxicity.

© Derrick Zellmann Photography