PSMA PET -- positron emission tomography using tracers that target prostate-specific membrane antigen -- has transformed the staging and management of prostate cancer. It can detect tumors and metastases far more precisely than older imaging methods, making it a cornerstone of modern prostate cancer diagnosis and treatment planning.
Despite this progress, standard visual inspection of PSMA PET images and conventional measurements like SUVmax (maximum standardized uptake value) capture only a fraction of the information contained in the scan. These simple metrics do not reliably predict tumor aggressiveness, whether cancer has spread locally, or who will experience cancer recurrence after treatment.
Radiomics offers a solution: rather than relying on what the human eye can see, it uses computer algorithms to extract hundreds of quantitative features from medical images -- measurements of texture, shape, intensity patterns, and spatial relationships between image voxels -- that reflect underlying tumor biology in ways invisible to radiologists.
When combined with artificial intelligence (AI) and machine learning (ML), these radiomic features can be used to build predictive models for tumor grade, local spread, and recurrence risk. This systematic review synthesizes 21 studies exploring this approach specifically for staging PSMA PET in newly diagnosed prostate cancer patients.
The authors searched three international databases -- PubMed, Scopus, and Web of Science -- using standardized keyword combinations involving prostate cancer, PSMA, PET imaging, and radiomics. Only original research articles on staging PSMA PET in patients with biopsy-proven, newly diagnosed prostate cancer who had not yet received any treatment were included.
From 166 initial records identified, duplicates and non-relevant articles were removed, leaving 21 studies covering a combined total of 2,497 patients. All studies were published after 2019, reflecting this field's very recent emergence. Three different PSMA-targeting radiotracers were used: gallium-68 PSMA-11 (57% of studies), fluorine-18 PSMA-1007 (33%), and fluorine-18 DCFPyL (10%).
Study quality was evaluated using the CASP (Critical Appraisal Skills Programme) checklist for diagnostic studies, and radiomics methodology quality was measured using the Radiomics Quality Score (RQS) developed by Lambin et al. -- a standardized tool that assesses adherence to best practices in radiomic analysis such as feature extraction standardization, validation approaches, and biological validation.
The 21 studies were grouped into three categories by clinical objective: predicting the ISUP grade group (a measure of tumor aggressiveness, 14 studies), predicting adverse pathologic features such as extracapsular extension or seminal vesicle involvement (9 studies), and predicting biochemical recurrence (BCR) after treatment (6 studies).
The most commonly studied application was predicting the ISUP grade group -- the standardized pathological score for prostate cancer aggressiveness -- before surgery, using only imaging data. This is clinically significant because biopsy-based grading can miss the most aggressive tumor regions due to sampling limitations.
Multiple studies demonstrated that radiomic features from PSMA PET could predict post-surgical ISUP grade group with AUCs typically in the 0.80 to 0.95 range. Importantly, combined PET plus MRI radiomic models outperformed PET-only or MRI-only models in several studies. One study (Solari et al.) showed that PET plus ADC (diffusion MRI) radiomic analysis outperformed prostate biopsy itself for predicting final pathological grade.
In one multiomics study (Ning et al.), a random forest model incorporating PSMA PET/MRI radiomics, genomics, and pathomics data achieved an AUC of 87% versus 75% for standard biopsy alone, with higher specificity (72% vs. 61%), suggesting fewer patients would be misclassified as lower risk. This has direct implications for treatment selection, particularly for active surveillance versus immediate treatment decisions.
A key challenge across studies was heterogeneity in which ISUP grade groups were used as thresholds for classification (some studies used ISUP 3+ vs. less than 3, others used 4+ vs. less than 4). This makes cross-study comparison difficult and highlights the need for consensus on clinical endpoints before these tools can be standardized.
Several studies tested whether PSMA PET radiomics could predict two critical signs of locally advanced disease: extracapsular extension (ECE) -- when cancer has spread beyond the prostate capsule -- and seminal vesicle invasion. Accurately predicting these before surgery is essential for planning the appropriate extent of resection and nerve preservation.
Cysouw et al. showed that machine learning models built on PET radiomic features from [18F]DCFPyL outperformed standard PET parameters for predicting ISUP grade, ECE, lymph node involvement, and distant metastases (p less than 0.01 for all). These results were subsequently validated in a second multicenter cohort by the same group, maintaining high prediction accuracy -- a rare and encouraging example of external validation in this field.
For seminal vesicle invasion, a random forest model that included a region of interest drawn over the seminal vesicles achieved an impressive AUC of 0.96, significantly outperforming visual radiologist assessment (AUC 0.70). For lymph node metastasis, a deep learning model showed 85% accuracy -- compared to 71% for five radiation oncologists making the same assessment from the same images.
One study (Pan et al.) directly compared PSMA PET/CT radiomic models with mpMRI radiomic models and a combined multimodal approach for predicting extracapsular extension. The mpMRI model was most accurate (AUC 0.85), and the multimodal model outperformed the PET model but did not add further improvement over mpMRI alone -- suggesting that MRI carries complementary and sometimes superior information for this specific task.
Biochemical recurrence (BCR) -- a rise in PSA after surgery or radiation that signals cancer has returned -- affects 20-50% of prostate cancer patients within five years of radical treatment and triggers further imaging, anxiety, and often systemic therapy with significant side effects. Predicting who is at high BCR risk before treatment would allow more personalized and potentially more aggressive initial management.
Papp et al. demonstrated that a machine learning model combining PSMA PET/MRI radiomics and clinical features achieved 89% accuracy for predicting BCR after radical prostatectomy, compared to 69% for standard PET metrics -- a 20-percentage-point improvement. It also predicted overall patient risk with 91% accuracy versus 70% for conventional parameters.
Li et al. built and externally validated a radiomic signature using [18F]F-PSMA-1007 PET to predict BCR-free survival. The model, using just three radiomic features selected by LASSO regression, achieved a C-index (a measure of discriminative ability for survival outcomes) of 76% in training and 71% in external validation. When clinical features were added, the combined model improved to C-indices of 81% and 78%, respectively.
Bian et al. used pre-surgical PSMA PET radiomics of periprostatic adipose tissue -- the fat surrounding the prostate -- to build a model predicting short-term post-surgical prognosis. This novel approach achieved AUCs of 85%, 77%, and 84% in training, internal, and external validation, respectively, illustrating that imaging information from tissue surrounding the tumor may carry additional prognostic value beyond the tumor itself.
The most significant limitation identified across the reviewed studies is the lack of external validation: only 5 of 21 studies tested their models on independent patient cohorts from different institutions. This is critical because radiomic features are sensitive to differences in PET scanner hardware, imaging protocols, and reconstruction parameters -- variations that are common across hospitals. A model that performs well at one center may fail at another if not specifically validated there.
The Radiomics Quality Score (RQS) evaluation revealed that while more than 57% of included studies exceeded the 40% quality threshold (better than the field average), none achieved outstanding scores. Issues identified included insufficient prospective design, lack of biological validation linking radiomic features to tumor molecular characteristics, and absent or inadequate external validation. These are long-standing systemic challenges in the radiomics field.
A promising methodological advancement used in some studies is ComBat harmonization -- an algorithm originally developed for genomics that corrects for scanner-related biases across multiple imaging sites. Studies applying ComBat showed improved generalizability across centers, suggesting this could be a key tool for enabling multicenter radiomic model deployment in the future.
The authors propose that PSMA PET radiomics could serve as a "digital biopsy" -- a non-invasive way to characterize the full biological landscape of the prostate tumor, overcoming the inevitable sampling limitations of needle biopsy, which misses aggressive regions in a meaningful percentage of cases. To reach clinical practice, the field needs prospective multicenter trials, consensus on imaging protocols, shared definitions of clinical endpoints, and rigorous external validation pipelines.