The standard path to prostate cancer diagnosis -- measuring blood PSA (prostate-specific antigen) levels followed by a 10-12 core biopsy -- has well-known limitations. Overdiagnosis and unnecessary biopsies remain common, and the biopsy procedure itself carries risks including infection, bleeding, and significant patient discomfort.
While multiparametric MRI (mpMRI) has improved pre-biopsy assessment, it has practical barriers: high cost, limited availability, and the fact that interpreting the standardized scoring system (PI-RADS) has a steep learning curve with substantial variation between radiologists. An affordable, widely available alternative would benefit many patients, especially in countries where MRI access is limited.
Ultrasound is already used to guide prostate biopsies, but standard B-mode ultrasound has limited ability to detect prostate cancer on its own. There is growing interest in extracting more information from ultrasound through radiomics -- the computational analysis of hundreds of quantitative image features that are invisible to the naked eye but may encode important information about tissue biology.
Radiomics refers to the automated extraction of large numbers of quantitative features from medical images -- features related to tissue texture, intensity patterns, shape, and spatial relationships that go far beyond what a radiologist perceives when visually inspecting an image. These mathematical features can then be fed into machine learning models to classify tissues or predict outcomes.
This study applied radiomics to two types of prostate ultrasound: standard B-mode ultrasound (grayscale structural imaging) and shear-wave elastography (SWE), a technology that measures tissue stiffness by tracking how sound waves propagate differently through hard versus soft tissue. Cancer tissue tends to be stiffer than normal prostate tissue, making SWE a complementary source of diagnostic information.
Combining radiomics features from both imaging modes into a single multiparametric model -- analogous to how MRI uses multiple acquisition types to characterize tissue -- was the central hypothesis of this study: that the combination would perform better than either mode alone.
The study enrolled 112 patients at a single hospital in Beijing who underwent transrectal ultrasound (TRUS) with B-mode and shear-wave elastography imaging, followed by a systematic 12-core biopsy plus additional targeted cores in suspicious areas (the "12+X" protocol). Biopsy pathology confirmed 54 patients had prostate cancer and 58 had benign lesions.
Radiologists manually outlined tumor regions on both B-mode and SWE images based on biopsy pathology records. From these regions, 1,218 features per imaging mode (2,436 total) were automatically extracted, spanning seven categories including first-order intensity statistics, shape features, and multiple texture matrix families (GLCM, GLRLM, GLSZM, GLDM).
Feature selection used LASSO regression -- a statistical technique that penalizes model complexity and automatically zeroes out uninformative features -- to reduce 2,436 features to just 20 key predictors for the combined multiparametric model. This produces a compact, weighted formula called the RAD-SCORE that summarizes the radiomics information into a single number for each patient.
Three models were built and compared using five-fold cross-validation: a clinical model (using age, PSA variants, PSA density, and other clinical variables), a radiomics model (using only the RAD-SCORE), and a combined clinical-radiomics model presented as a visual nomogram. Performance was evaluated using the area under the receiver operating characteristic curve (AUC) and decision curve analysis.
Among the radiomics models, adding shear-wave elastography improved accuracy over B-mode alone: the SWE radiomics model achieved an AUC of 0.80 compared to 0.74 for B-mode alone. The combined multiparametric RAD-SCORE (integrating both modes) reached an AUC of 0.85 -- meaningfully better than either single-mode approach.
The clinical model (using age, PSA density, and other clinical factors) achieved an AUC of 0.84, similar to the multiparametric radiomics model. However, the combined clinical-radiomics model outperformed both, achieving an AUC of 0.90 in the validation group -- demonstrating that radiomics and clinical variables capture complementary information that together provides better discrimination than either source alone.
Decision curve analysis confirmed the clinical advantage of the combined model: it provided higher net benefit across a wide range of risk thresholds compared to either the clinical model or radiomics model alone. This means that in practice, using the combined nomogram would result in better clinical decisions (fewer unnecessary biopsies and fewer missed cancers) compared to using clinical factors or radiomics independently.
Analysis of the key selected features provides biological insight into why the model works. Cancer tissue showed lower mean gray-level intensity in B-mode ultrasound than benign tissue -- consistent with the known appearance of prostate cancer as hypoechoic (darker) lesions on ultrasound. Radiomics quantifies this as an objective, reproducible number rather than a subjective visual impression.
Benign lesions (primarily benign prostatic hyperplasia) showed lower gray-level correlation values and higher spatial heterogeneity in pixel intensity patterns compared to cancer. This likely reflects the biological heterogeneity of BPH -- which contains mixtures of gland cells, smooth muscle, and inflammatory cells -- versus the more uniform cellular proliferation pattern of prostate cancer.
For shear-wave elastography, cancer lesions showed higher mean stiffness values (Mean@SWE) -- confirming that prostate cancer tissue is harder than benign tissue. This stiffening occurs because cancer cells infiltrate the interstitial tissue and trigger a desmoplastic reaction (hardening of surrounding connective tissue) even before structural changes become visible on conventional imaging.
The final combined model was presented as a nomogram -- a graphical tool that allows a clinician to estimate the probability of prostate cancer by drawing lines across scales for patient age, PSA density, and the multiparametric RAD-SCORE. The nomogram converts complex statistical calculations into an intuitive, easy-to-use visual format suitable for routine clinical use.
The independent predictors in the final model were patient age and PSA density (PSA adjusted for prostate volume) -- confirming prior evidence that PSA density provides more individualized information than raw PSA alone, since it accounts for the fact that a larger prostate naturally produces more PSA regardless of cancer status.
Unlike MRI-based tools which require expensive equipment and specialized radiologist training, this ultrasound-based approach uses equipment already present at most facilities performing prostate biopsies. A nomogram incorporating ultrasound radiomics could be integrated into the standard pre-biopsy workflow with minimal additional burden, potentially reducing unnecessary biopsies and improving risk stratification at the point of care.
The study has important limitations including its retrospective design at a single institution with only 112 patients. Although five-fold cross-validation was used internally, external validation in a larger, multi-center cohort is needed before the model can be recommended for clinical use. The small sample size also meant that differences in RAD-SCORE between different Gleason grades (levels of cancer aggressiveness) could not be reliably detected.
The model does not currently distinguish between the peripheral zone and transition zone of the prostate, which have different cancer rates and imaging appearances. Future work incorporating zone-specific analysis may further improve accuracy. Including contrast-enhanced ultrasound (CEUS) as a third imaging mode could also add complementary perfusion information to the multiparametric model.
False positives can arise from conditions like prostatitis or benign prostatic hyperplasia that also stiffen tissue and change ultrasound appearance. Immunohistochemical analysis of false-positive cases may help identify specific imaging features that distinguish these conditions from cancer, further refining the model's performance.