At University Hospital Southampton, researchers have built a machine-learning tool that can predict how much radiation a tumor will receive during advanced prostate cancer therapy—before a single dose is given. The breakthrough could transform treatment planning for patients with metastatic castration-resistant prostate cancer, a particularly aggressive form of the disease.
Currently, doctors don't know the precise radiation dose delivered to a patient's tumors and surrounding organs until after treatment has already begun. That information typically comes from time-consuming, resource-intensive post-therapy imaging scans. Amit Nautiyal, a scientist and National Institute for Health and Care Research fellow at University Hospital Southampton and the University of Southampton, realized that hospitals were already sitting on useful data: routine pre-therapy PET/CT scans performed before ¹⁷⁷Lu-PSMA radiopharmaceutical therapy. "18F-PSMA PET/CT is already routinely performed and widely available in prostate cancer patients, but its potential to predict treatment radiation dose has not previously been explored," Nautiyal explained.
In their proof-of-concept study, Nautiyal's team worked with nine patients with metastatic castration-resistant prostate cancer who were scheduled for ¹⁷⁷Lu-PSMA therapy. These patients contributed 57 tumors, 36 salivary glands, and 18 kidneys to the analysis—providing enough data to train a machine-learning model. The researchers fed the model with three types of information from pre-therapy scans: uptake-based PET metrics (how much the tumor absorbed the tracer), radiomics features (detailed imaging patterns), and clinical biomarkers from blood work and patient history. The model then predicted the radiation dose each tumor and healthy organ would receive.
When researchers compared the model's predictions against the actual dosimetry calculated after the first cycle of therapy, the results were encouraging. The machine-learning approach demonstrated a promising ability to forecast both tumor and organ absorbed doses by combining multiple data sources while accounting for how individual patients' bodies respond differently.
Why does this matter? Knowing the expected dose beforehand could help doctors make smarter decisions about treatment. Some patients might receive a lower dose based on predictions—protecting their health. Others might be identified upfront as poor candidates for the therapy, sparing them from a treatment unlikely to help. This represents a shift in how imaging works: instead of simply diagnosing disease, imaging data actively guides personalized treatment decisions.
"If validated in larger studies, this approach may improve patient selection and support better decision-making during pre-treatment assessment, helping to optimize ¹⁷⁷Lu-PSMA therapy for individual patients," Nautiyal noted. The team presented their findings at the Society of Nuclear Medicine and Molecular Imaging 2026 Annual Meeting.
This is just the beginning. Nautiyal and colleagues are embarking on a planned five-year program to collect more patient data and build a more robust, validated model. The next phase will involve larger, multi-center cohorts across different hospitals, refining the predictions and testing the tool's accuracy in independent groups of patients. Once fully validated, the tool could be deployed in clinical practice to support treatment decisions for patients facing one of prostate cancer's most dangerous forms.