Metastatic prostate cancer (mPCa) dramatically worsens survival: the 5-year survival rate drops from nearly 100% for localized disease to just 34.1% when cancer has spread to distant sites. Bone is the most common metastatic site, accounting for up to 90% of metastatic cases. Early and accurate detection of metastatic spread is critical for initiating the most effective treatments, including androgen deprivation therapy (ADT), which has been shown to reduce mortality and skeletal fractures when started early.
PSMA PET scans -- positron emission tomography using radiotracers that target prostate-specific membrane antigen -- have transformed how prostate cancer is staged. PSMA is a protein highly expressed on prostate cancer cells, making it an ideal imaging target. The landmark ProPSMA trial demonstrated that PSMA PET-CT outperforms conventional CT plus bone scan for cancer staging, and two PSMA tracers now carry FDA approval in the US.
Despite its high diagnostic accuracy, PSMA PET reporting suffers from significant inter-reader variability -- different specialists reading the same scan may reach different conclusions. Standardization efforts such as the PROMISE criteria and PSMA-RADS scoring system have improved consistency, but these tools are labor-intensive and time-consuming for busy clinicians.
Artificial intelligence offers a path forward: by automating lesion detection, classification, and quantification, AI could reduce the workload on nuclear medicine specialists, minimize subjective variability, and potentially identify patterns in scans that human readers might miss. This systematic review synthesizes available evidence on AI's ability to enhance PSMA PET interpretation for metastatic disease.
The review was conducted according to PRISMA guidelines and registered on PROSPERO (ID CRD42023456044), the international registry for systematic reviews. A comprehensive literature search was performed across three major databases -- Medline, Embase, and Scopus -- from their inception through July 2023, using search terms covering artificial intelligence, machine learning, deep learning, prostate cancer, and PSMA PET.
The search returned 249 articles. After removing 80 duplicates and screening 169 titles and abstracts, 28 studies were reviewed in full. Following full-text review, 11 studies met the eligibility criteria. Studies were included if they used AI to evaluate metastatic prostate cancer or lymph node involvement specifically on PSMA PET scans. Studies using non-PSMA tracers, evaluating only intra-prostatic lesions, or presented as case reports or reviews were excluded.
Study quality was assessed using the STREAM-URO 26-item checklist designed specifically for machine learning studies in urology, and risk of bias was evaluated with the PROBAST tool. Of the 11 studies, 10 showed low overall risk of bias. The mean STREAM-URO score was 21 out of 28, with the main weaknesses being incomplete reporting of patient demographics and missing exclusion criteria.
The heterogeneity of included studies -- varying AI algorithms, PSMA tracers, patient populations, and outcome measures -- precluded a formal meta-analysis. Results are therefore presented descriptively, organized by the application category of the AI model.
AI demonstrated high accuracy for detecting metastatic disease. Two machine learning studies reported AUC values of 0.98 with sensitivities of 97% and 94% for identifying suspicious uptake outside the prostate. A convolutional neural network (CNN) separately achieved 80.4% average precision for suspicious uptake identification. These results are comparable to or better than expert nuclear medicine physicians in some cases.
For lymph node detection, the FDA-approved aPROMISE deep learning software achieved 91.5% sensitivity for regional lymph nodes and 90.6% for all lymph nodes combined. A separate CNN achieved 81% agreement with expert reviewers for regional lymph node identification. These figures are clinically important because lymph node status directly determines treatment strategy, particularly for decisions about lymph node dissection or extended field radiation.
For bone metastasis detection, sensitivities ranged from 62% to 86.7% across different AI systems. Notably, in one study the CNN matched or slightly exceeded the nuclear medicine physician's sensitivity (62% vs. 59%), though with a lower positive predictive value (40.5% vs. 58.7%). AI also demonstrated the ability to differentiate active bone metastases from post-treatment sclerotic bone changes -- a clinically difficult distinction -- with 73.5% accuracy using a weighted k-nearest-neighbor algorithm.
AI showed additional capabilities beyond basic detection, including standardizing reporting by automatically assigning PSMA-RADS categories (with patient-level AUROC of 0.9) and automatically segmenting anatomical regions with Dice scores comparable to expert radiologists (0.88-0.97 across different organs and tissue types).
Beyond detection, AI demonstrated ability to quantify tumour burden -- a measure of how much cancer is present across the body. Two studies trained CNNs to measure total lesion volume (TLV) and total lesion uptake (TLU), both important prognostic indicators. The CNN results correlated strongly with nuclear medicine physician calculations (Spearman R = 0.53 to 0.83), validating their reliability.
In a prospective study by Kendrick et al. -- the only prospective study in this review -- a CNN extracted TLV and TLU from whole-body PSMA PET scans with 94.5% overall accuracy. Crucially, both TLV and TLU calculated by the CNN correlated significantly with patient overall survival (p < 0.005), meaning the AI-derived tumour burden measurement has direct prognostic value.
For treatment response monitoring, the aPROMISE platform was used to quantify changes in PSMA signal before and after treatment with surgery, radiotherapy, or ADT. Changes in PSMA score correlated significantly with PSA reductions post-treatment for nodal disease, suggesting AI-computed PSMA scoring could serve as an objective marker of treatment efficacy.
One study went further, using a machine learning model trained on baseline PSMA PET radiomics and clinical parameters to predict response to 177Lu-PSMA -- a radioligand therapy for metastatic prostate cancer -- with an AUC of 80%, 75% sensitivity, and 75% specificity. This demonstrates AI's potential for pre-treatment response prediction, which could guide which patients are most likely to benefit from this costly and selective treatment.
The most immediately actionable AI application in PSMA PET reporting is automated organ segmentation and anatomical allocation. This is a labor-intensive prerequisite step before any lesion classification can occur. AI performs this task with accuracy comparable to experts (Dice scores 0.88-0.97), reducing the time burden on nuclear medicine specialists significantly.
The authors recommend that AI tools should be used as decision support rather than autonomous reporting systems. Given the variable sensitivity and relatively low positive predictive value in some studies (PPV ranging from 39.2% to 66.8%), an experienced nuclear medicine physician should review and proofread all AI-generated reports before clinical use. However, AI can serve as an educational tool for trainees, helping them identify positive sites during supervised learning.
An important emerging clinical scenario is managing patients with positive regional lymph nodes on PSMA PET who had negative conventional staging. If AI can precisely characterize the probability that a PSMA-avid lymph node is truly malignant, it could help clinicians avoid unnecessary lymph node dissections or extended radiation fields for low-probability nodes, reducing treatment morbidity.
As PSMA PET becomes the standard pre-treatment staging modality, AI could bridge the gap between the new imaging capabilities and the existing treatment guidelines, which were developed in the conventional staging era. AI's ability to extract precise quantitative data from PSMA scans could support the creation of new evidence-based criteria for treatment decisions in PSMA-staged patients.
A key limitation across studies is the wide variation in sensitivity (62-97%) for metastasis detection. Possible explanations include differences in training data (some models may perform better on high-volume disease and miss small lesions), different PSMA tracers (68Ga-PSMA vs. 18F-PSMA have different pharmacokinetics), and variations in scanner hardware and acquisition protocols between institutions.
Most studies were retrospective with small sample sizes, and many failed to report basic patient demographics such as age and PSA level. Only one of the 11 studies was prospective. Without large, diverse, prospective datasets, the true generalizability of these AI models to real-world clinical settings remains uncertain.
Only one study incorporated clinical parameters (age, PSA, Gleason score) alongside imaging features in the AI model. This is a missed opportunity, as combined models integrating imaging and clinical data consistently outperform imaging-only models in other cancer contexts. Future studies should explore whether adding clinical parameters improves PSMA PET AI model performance.
The black box nature of many AI algorithms -- where the model's reasoning process is not transparent to the clinician -- presents a practical barrier to clinical adoption. As the field moves toward explainable AI, models that can highlight which scan regions drove a prediction will be more trustworthy and adoptable. Regulatory approval pathways also need to evolve alongside the technology to facilitate safe clinical integration.
This systematic review demonstrates that AI can detect lymph node involvement and metastatic prostate cancer on PSMA PET scans with high accuracy (AUC up to 98%) and sensitivity ranging from 62-97%. Additional documented capabilities include tumour burden quantification, treatment response monitoring, bone lesion differentiation, and automated reporting standardization.
The FDA-approved aPROMISE platform represents the most clinically mature AI application for PSMA PET, already validated in multi-center settings and capable of automated organ segmentation, PSMA uptake quantification, and standardized staging output. It provides a concrete example of AI transitioning from research tool to clinical product in nuclear medicine.
Future research priorities include large-scale prospective multi-center validation studies, AI models that combine imaging radiomics with clinical and genomic parameters, models specifically trained to detect small lymph nodes (currently a major PSMA PET limitation -- 91% of missed metastatic nodes are under 5mm), and explainable AI approaches that make model decisions transparent to clinicians.
Longer-term research should also investigate whether AI-enhanced PSMA PET detection translates into improved long-term oncological outcomes -- demonstrating not just diagnostic superiority but actual patient benefit in terms of survival and quality of life, which is ultimately the measure that matters most in clinical medicine.