Metastatic prostate cancer (mPCa) carries a dramatically worse prognosis than localized disease. While early-stage prostate cancer has a 5-year survival rate above 90%, that figure drops below 40% once the cancer spreads and falls under 10% when bone metastases are present.
Standard diagnostic tools -- including the PSA test, digital rectal exam (DRE), and TRUS-guided biopsy -- have significant limitations. PSA is non-specific and can be elevated in benign conditions. DRE has a high false-positive rate, and biopsy samples only small portions of the tumor, missing the spatial complexity of the disease.
Prostate cancer is highly heterogeneous, meaning lesions vary widely in aggressiveness from site to site and over time. This heterogeneity makes it difficult to characterize the full extent of disease from a single biopsy and drives the need for whole-body, non-invasive assessment tools.
Medical imaging technologies -- including mpMRI, PET, CT, and bone scintigraphy -- are already routinely used in cancer management, but radiologists typically assess them qualitatively. A more systematic, mathematical approach to extracting information from these scans could unlock significant diagnostic value.
Radiomics is the large-scale extraction of quantitative mathematical features from medical images to characterize tumor biology. The core idea is that patterns in imaging data -- invisible to the naked eye -- reflect underlying tumor characteristics that can predict behavior and guide treatment.
Unlike a radiologist's subjective interpretation, radiomics produces objective, reproducible numerical measurements. These quantitative imaging biomarkers can overcome inter-observer variability and may detect differences between tumor subtypes that are imperceptible to human review.
Importantly, radiomics is non-invasive. Because imaging scans are already part of standard clinical care for cancer patients, incorporating radiomics analysis into the workflow does not require additional procedures. It derives new value from existing data.
The Radiomics Quality Score (RQS) was developed to evaluate the scientific rigor of radiomics studies. It awards or deducts points based on 16 criteria -- including whether protocols are documented, whether external validation was done, and whether the study was prospective. The maximum score is 36 points.
The radiomics workflow begins with image acquisition, where parameters like slice thickness, pixel spacing, reconstruction algorithm, and scanner vendor can all affect the final feature values. Comprehensive documentation of these parameters is essential for reproducibility across institutions.
The next critical step is region of interest (ROI) segmentation -- drawing boundaries around the tumor in the image. This can be done manually by clinicians, semi-automatically, or by fully automated algorithms. Manual segmentation introduces inter-observer variability and is time-consuming; automated methods can reduce bias but require validation.
After segmentation, hundreds to thousands of hand-crafted features are extracted -- including shape features (tumor compactness, sphericity), first-order statistics (intensity distributions), and texture features derived from matrices like the grey-level co-occurrence matrix (GLCM). These capture both spatial structure and intensity patterns within the tumor.
An alternative to hand-crafted features is deep learning, where convolutional neural networks (CNNs) automatically learn image patterns without human-defined feature engineering. While powerful, these learned features are difficult to interpret, creating a tradeoff between predictive performance and clinical transparency.
Feature selection trims redundant or non-informative variables before model building. Techniques range from statistical filtering to principal component analysis (PCA) to regularized regression methods like LASSO. The final predictive models use algorithms such as logistic regression, support vector machines, random forests, or neural networks.
Several studies used mpMRI radiomics features to predict bone metastases before they occur. One model combining features from T2-weighted and dynamic contrast-enhanced MRI with clinical variables (age, Gleason score, PSA) achieved an AUC of 0.895 in predicting future bone metastases -- significantly outperforming Gleason score alone (AUC = 0.731).
PET imaging with tracers targeting prostate-specific membrane antigen (PSMA) proved highly effective for detecting existing metastases. A study of 76 patients using 18F-DCFPyL PET achieved AUC values of 0.86 for both lymph node involvement and distant metastases when radiomics features were analyzed with a random forest model.
A study combining traditional radiomics with deep learning transfer features from five pre-trained neural networks achieved AUC = 0.76 in predicting pelvic lymph node metastases in an external validation set of 50 patients. The combined model outperformed both the Memorial Sloan Kettering nomogram and the Briganti score for identifying patients who could safely avoid extended pelvic lymph node dissection.
Multi-modal analysis consistently improved performance. Using both ADC maps and T2-weighted MRI features together for lymph node prediction yielded accuracy of 0.90 -- better than either modality alone (ADC: 0.86, T2w: 0.84). These results support combining imaging data from multiple sources to capture complementary information about tumor biology.
Radiomics can assess how well metastatic lesions respond to treatment. In patients undergoing 177Lu-PSMA radioligand therapy -- a targeted treatment for advanced prostate cancer -- texture features like entropy and homogeneity of bone lesions correlated with PSA changes following therapy, suggesting that more heterogeneous lesions responded better.
A study using non-contrast CT images demonstrated that radiomics could distinguish between sclerotic bone lesions that had fully responded to treatment versus those still harboring active metastases. This is clinically significant because responded sclerotic lesions retain their imaging appearance even after cancer is cleared, making visual distinction unreliable.
Deep learning models applied to bone scintigraphy images achieved strong results in diagnosing the presence or absence of bone metastases. CNN architectures trained on hundreds of patient scans achieved overall accuracy of 91-97%, with one lightweight network matching state-of-the-art performance while requiring fewer computational resources.
For PET-based lesion classification, an Extra Trees machine learning model applied to 68Ga-PSMA PET/CT features classified benign versus malignant hotspots with AUC = 0.98. A dual-CNN system for automated lymph node metastasis detection found that the number of automatically detected lesions was significantly associated with cancer-specific survival.
The RQS assessment of the 17 traditional radiomics studies in this review revealed a mean score of only 23% (out of 100%), with a range from 0% to 52.8%. This is consistent with the low methodological quality documented across radiomics research in other cancer types.
Only 3 of the 17 studies (17.6%) conducted external validation of their models -- a critical requirement for clinical applicability. Only 4 studies (23.5%) were prospective in design, meaning most were based on retrospective data with its associated selection bias.
Additional weaknesses included a near-complete absence of biological correlation discussion (0/17 studies), no cost-effectiveness analysis (0/17), and 94.1% of studies failing to follow open science principles by sharing code, data, or segmentations publicly.
Deep learning studies had similar quality issues. None were prospective, and only a single deep radiomics study validated results in an external dataset. These limitations mean that although individual model results are often impressive, they cannot yet be considered generalizable for real-world clinical use.
Reproducibility is the core challenge for radiomics. Features extracted from the same image can differ significantly depending on scanner manufacturer, acquisition protocol, reconstruction algorithm, and post-processing choices. This makes it difficult to apply a model trained at one institution to patients scanned at another.
Harmonisation techniques like ComBat statistical correction can partially adjust for scanner differences in retrospective datasets. However, the most reliable solution is prospective standardization -- agreeing on imaging acquisition protocols before data collection begins and validating results across multiple institutions.
Manual tumor segmentation introduces significant variability. Moving toward automated or semi-automated segmentation algorithms would reduce this source of bias, and CNN-based models that operate without needing a pre-drawn ROI may bypass this limitation entirely.
The combination of radiomics with clinical covariates -- such as PSA level, Gleason score, and patient age -- consistently improves model performance. Future models should integrate multi-modal data from imaging, blood biomarkers, and molecular profiling to capture the full biological complexity of metastatic prostate cancer.
The review authors recommend that future radiomics studies in metastatic prostate cancer adopt prospective study designs whenever possible, with pre-specified imaging protocols and planned external validation cohorts from other institutions.
Studies should prioritize feature reproducibility analysis -- including multiple segmentations by different observers and phantom studies to assess sensitivity to scanner parameters -- before building and reporting predictive models.
Radiomics software and analysis pipelines should follow open science principles: sharing code, segmentation files, and extracted features publicly so that results can be replicated, challenged, and improved by the broader research community.
Despite current limitations, radiomics holds genuine potential to transform the management of metastatic prostate cancer. By incorporating quantitative imaging biomarkers into clinical workflows, physicians could achieve more precise patient stratification, earlier detection of disease spread, and better-informed treatment decisions -- all without additional invasive procedures.