AI in radiopharmaceutical therapy: powerful tools, but no proof yet they change treatment decisions

Artificial intelligence is being woven into nearly every step of radiopharmaceutical therapy — the fast-growing family of cancer treatments in which doctors image a molecular target, then inject a radioactive drug designed to destroy it — but a sweeping new evidence map from the University of Miami Miller School of Medicine concludes that, so far, almost none of it has been shown to actually change what happens to a patient. The review, published in Frontiers in Nuclear Medicine, combed four major databases and mapped 73 peer-reviewed journal reports plus eight conference abstracts in which a machine-learning or deep-learning model was trained and evaluated on a task feeding a theranostic decision. The verdict is nuanced: AI is technically impressive and increasingly useful for measurement and workflow, but the evidence for treatment-changing use remains strikingly thin.

The clinical stakes are high because radiopharmaceutical therapy (RPT) is conceptually precise yet operationally crude. Prostate-specific membrane antigen (PSMA)-targeted radioligand therapy for prostate cancer, somatostatin-receptor peptide receptor radionuclide therapy (PRRT) for neuroendocrine tumors, radioiodine therapy for thyroid cancer, and yttrium-90 radioembolization for liver tumors all embody the theranostic principle: image the target, treat the target, quantify the biodistribution, absorbed dose, and response. Yet despite this conceptual elegance, many workflows still administer fixed or empirically chosen activities, and the supporting physics — labor-intensive segmentation, serial imaging, activity quantification, time-activity modeling, and absorbed-dose reporting — remains slow and inconsistent across centers.

AI is attractive precisely because the RPT workflow is data-rich and quantitatively constrained. Patients undergo baseline PET/CT or SPECT/CT, post-therapy quantitative imaging, serial laboratory testing, symptom and toxicity monitoring, and repeated treatment cycles. These data can be linked to activity concentration, time-integrated activity, absorbed dose, effective half-life, tumor burden, and clinical measures of organ reserve. Learned models may reduce the segmentation burden, standardize tumor-burden quantification, infer kinetic parameters, approximate computationally expensive dose calculations, and estimate response or toxicity risk. The central clinical question, the authors argue, is not whether AI can produce a high-performing model in a retrospective dataset, but whether an AI-supported output can safely and reproducibly improve treatment decisions — whether to treat, whether to intensify or de-escalate activity, whether to delay or discontinue a cycle, whether the predicted organ dose is acceptable, and whether early response suggests continued benefit.

The mapped evidence set is revealing in its composition. Among the 73 full journal reports, 17 were PSMA-related, 16 addressed SSTR/PRRT, 15 radioiodine, 14 yttrium-90 radioembolization, and 11 were cross-platform or emerging-target studies. Sixteen addressed segmentation or quantitative tumor burden, 32 addressed selection, response, prognosis, toxicity, or safety and logistics, and 25 addressed registration, preprocessing, or dosimetry. But the validation picture is sobering: 51 reports were retrospective, only one was an explicitly prospective clinical or technical evaluation, nine were technical, synthetic, or phantom evaluations, and the temporal design was unclear or conflicting in 12. Only seven reported an external-type held-out evaluation, and just four clearly held out an entire institution. One report described explicit calibration, one propagated task-level uncertainty, and none evaluated a prospective AI-guided treatment policy.

The technical highlights are nonetheless substantial. Zhao and colleagues trained a deep neural network to detect and segment PSMA-avid lesions across 130 training patients and 63 patient-separated test patients, achieving F1 scores of 0.99 for bone lesions and 0.92 for lymph-node lesions. Perret and colleagues evaluated nnU-Net lesion segmentation on paired gallium-68 PSMA PET/CT and first-cycle lutetium-177 PSMA SPECT/CT in 73 patients, training at one center and testing at a second — a genuine held-out-center validation yielding external Dice scores of 0.76 to 0.77 for PET and 0.78 for PET-guided SPECT, though PET volume errors were highly unstable and activity or absorbed-dose consequences were never tested. In PRRT dosimetry, Nakaichi and colleagues reported that deep-learning segmentation cut the median total workflow time from 54.3 to 47.0 minutes, a modest but statistically significant acceleration — though the reference contour was the staff-corrected AI output itself, meaning the study validated supervised workflow support rather than autonomous segmentation.

Dosimetry — the quantitative heart of personalized RPT — shows both the promise and the peril. Mansouri and colleagues used a hybrid transformer model to predict lutetium-177 DOTATATE voxel dose maps against Monte Carlo references in 22 patients, reporting a voxel relative absolute error of 5.28 percent with patient-grouped cross-validation. Mansouri and colleagues also applied Swin UNETR deep learning to attenuation and scatter correction of technetium-99m MAA SPECT in 222 patients toward yttrium-90 radioembolization quantification, with a blind internal holdout showing voxel mean dose error of at most 0.003 Gy. Jalilifar and colleagues trained a multilayer perceptron to estimate patient-specific iodine-131 dosimetry from a single oblique planar view, with organ-specific differences from Monte Carlo reference ranging from −14.8 to +11.2 percent in a 23-patient external cohort. Georgiou and colleagues trained a feed-forward neural network to estimate the MIRD maximum permissible iodine-131 activity from reduced-time-point imaging, with a nonsignificant paired difference against full five-time-point calculations — but the authors stress that a nonsignificant difference does not establish equivalence without a prespecified clinical margin.

On the decision side, the picture is more cautious. In PSMA therapy, Pan and colleagues developed a dual-tracer radiomics nomogram to estimate PSMA-negative but FDG-positive disease — a biological mismatch that predicts radioligand therapy failure — achieving an AUC of 0.83 in development, with decision analysis suggesting a 5 percent threshold could avoid 21 percent of FDG PET/CT examinations while missing only 2 percent of mismatch cases. But a separate cohort tested only the prognostic associations, not the primary classification performance. In PRRT, Baur and colleagues used multimodal deep learning fusing PET/CT and laboratory variables to predict progression-free survival in 116 patients, achieving an AUC of 0.72 with repeated cross-validation — retrospective, single-center, internally validated evidence. In radioiodine therapy, Luo and colleagues used LightGBM with SHAP interpretability in 950 patients to predict non-excellent response, achieving a test AUC of 0.896 versus 0.842 for logistic regression — but because the model incorporated administered activity and post-therapy scan variables, it informs post-treatment response stratification rather than pretreatment selection. Popovic and colleagues trained an artificial neural network to imitate an experienced physician’s radioiodine activity recommendations, achieving 95.71 percent test accuracy and weighted kappa of 0.96 — expert-decision imitation, not outcome or treatment-benefit validation.

The authors organize these gaps into an eight-domain evidence-to-decision synthesis covering quantitative fidelity, reference-standard appropriateness, external or prospective validation, clinical-endpoint relevance, dose-response or dose-toxicity linkage, workflow and human oversight, uncertainty and quality assurance, and decision impact. This is explicitly a narrative organizing aid, not a validated score or adoption standard. The domains were derived by mapping recurrent gaps across the retained evidence and comparing them with established reporting frameworks including RELAINCE, RQS 2.0, CLAIM, TRIPOD+AI, PROBAST+AI, and DECIDE-AI. The RPT-specific emphasis is the separation of dosimetric fidelity, biological linkage, and treatment action — three things that are often conflated in the literature but must be established independently before any AI output can justify modifying a prescription.

Four inferential targets recur throughout the review, and distinguishing them is essential. Prognostic models estimate risk under observed care; response models estimate an endpoint after treatment; treatment-effect models estimate outcomes under specified alternatives; and treatment policies map patient states to actions. Only the latter two can directly justify choosing or withholding treatment. The authors note that target trial emulation, causal machine learning, and individualized treatment-effect estimation remain hypothesis-generating unless timing, eligibility, confounding, positivity, censoring, overlap, calibration, and decision impact are addressed — and even then they require prospective evaluation of a prespecified management rule. Dataset limitations compound the problem: lesion-level random splits can leak correlated observations across training and testing when multiple lesions arise from one patient, and reused or overlapping institutional cohorts can make apparently independent studies less independent than their citations suggest.

The research roadmap that emerges is pragmatic. The highest-yield infrastructure goal is a curated multicenter dataset linking diagnostic imaging, post-therapy quantitative imaging, laboratory values, administered activity, absorbed-dose estimates, response and toxicity endpoints, and longitudinal outcomes — with shared acquisition, reconstruction, calibration, segmentation, and dosimetry metadata so models can be compared under common rules. Multicenter dose-response consortia should pool patient-, organ-, lesion-, and voxel-level dosimetry with standardized clinical outcomes. Future trials should test prespecified adaptive strategies — response-adapted continuation or discontinuation, toxicity-aware delay, activity escalation or de-escalation — only when target expression, organ reserve, uncertainty bounds, and protocol-defined safety constraints support these actions. The near-term opportunity, the authors conclude, is human-supervised decision support rather than fully autonomous adaptation: a useful system should display its inputs, uncertainty, reference standards, validation settings, and proposed clinical actions, and flag cases outside its development domain. Treatment-changing use will require standardized quantitative imaging, appropriate dosimetry comparators, multicenter and prospective validation, calibrated uncertainty, clinically established dose-response or dose-toxicity relationships, and a prespecified evaluation of decision impact — a demanding but achievable sequence that would finally convert technical feasibility into clinical benefit.

Subject of Research: Artificial intelligence in oncologic theranostics and radiopharmaceutical therapy

Article Title: Artificial intelligence across oncologic theranostics: evidence for patient stratification, dosimetry, and adaptive radiopharmaceutical therapy

Article References: Banala, M., Ezuddin, S., Foley, M., & Kuker, R. (2026). Artificial intelligence across oncologic theranostics: evidence for patient stratification, dosimetry, and adaptive radiopharmaceutical therapy. Frontiers in Nuclear Medicine, 6, Article 1935247. https://doi.org/10.3389/fnume.2026.1935247

Image Credits: AI Generated

DOI: 10.3389/fnume.2026.1935247

Keywords: artificial intelligence, radiopharmaceutical therapy, theranostics, dosimetry, PSMA, PRRT, radioiodine therapy, nuclear medicine, machine learning, cancer treatment, Artificial, intelligence