Prostate cancer is the fourth most commonly diagnosed cancer in men and the fifth leading cause of male cancer death worldwide. Its detection and risk assessment are particularly challenging because prostate cancer ranges from extremely slow-growing tumors that will never cause harm to aggressive cancers requiring immediate treatment.
Traditional imaging using transrectal ultrasound (TRUS) has limited sensitivity and specificity of only 40 to 50% for detecting prostate cancer, making it an unreliable screening tool. PSA blood testing is widely used but cannot reliably distinguish aggressive from indolent disease.
Multiparametric MRI (mpMRI) has transformed this situation by combining three imaging sequences: T2-weighted imaging for anatomical detail, diffusion-weighted imaging (DWI) for cellular density, and dynamic contrast-enhanced (DCE) imaging for blood vessel activity. This combination achieves sensitivity of 72% and specificity of 81% for detecting clinically significant cancer.
The PI-RADS scoring system (version 2.1, published 2019) standardizes how radiologists report mpMRI findings on a 1-to-5 scale from highly unlikely to highly likely to be cancer. However, PI-RADS 3 and 4 lesions remain clinically ambiguous, representing a moderate-to-high risk that still requires biopsy to resolve, leaving significant room for improvement.
Radiomics is a process that extracts large numbers of quantitative numerical features from medical images, converting visual information into analyzable data. The core idea is that tumors contain spatial and textural patterns encoded in their pixel values that carry biological information invisible to the human eye.
Radiomics exists within a hierarchy of artificial intelligence approaches. Machine learning uses manually selected features to train statistical classifiers. Representation learning automatically discovers useful features from raw data. Deep learning, specifically convolutional neural networks, learns hierarchical feature representations ranging from simple edges and textures to complex shapes and tissue patterns.
The complete radiomics workflow involves five sequential steps: segmentation (defining the tumor region), image processing (standardizing appearance), feature extraction (computing numerical descriptors), feature selection (identifying the most informative subset), and development of a predictive model using machine learning algorithms.
This review summarizes the current state of radiomics research applied specifically to prostate cancer on mpMRI, covering four application areas: cancer detection, diagnosis, Gleason grade prediction, PI-RADS score refinement, and treatment outcome prediction.
Segmentation defines the region or volume of tissue from which features will be extracted. Manual segmentation is time-consuming and suffers from intra- and inter-observer variability. Automated methods using atlas-based or deep learning approaches are preferred because they are faster and more reproducible, though they work better for homogeneous than heterogeneous lesions.
Image processing standardizes images before feature extraction by correcting intensity non-uniformities, normalizing gray-level ranges, and resampling to uniform voxel spacing. This step is essential because scanner-to-scanner differences in image acquisition would otherwise make features incomparable across patients or institutions.
Feature extraction computes two broad types of descriptors. First-order statistics summarize the distribution of individual pixel intensities (mean, entropy, energy). Second-order statistics, derived from Gray-Level Co-occurrence Matrices (GLCM) and Gray-Level Run-Length Matrices (GLRLM), capture how pixel intensities relate to each other spatially, quantifying texture complexity and heterogeneity within the tumor.
Feature selection reduces hundreds of extracted features to a compact set of non-redundant, reproducible predictors using methods like cluster analysis and principal component analysis. This step prevents overfitting, which occurs when a model memorizes noise in the training data rather than learning genuine biological patterns. The selected features then train a classifier such as a support vector machine, random forest, or neural network.
Early computer-aided detection (CAD) systems using radiomics demonstrated that quantitative MRI analysis could systematically improve detection sensitivity. Reported improvements range from 74% to 100% sensitivity with 43% to 93% specificity using 1.5 Tesla MRI scanners, with performance depending heavily on the imaging sequences included and the classifier used.
Giannini et al. created a two-stage CAD system that first produced a color-coded probability map of the entire prostate, assigning each voxel a probability of being cancerous based on multiparametric features, then used this map to segment candidate tumor regions for clinical review. This approach achieved AUC values of 0.83 to 0.98 and demonstrated that fully automated detection was feasible.
A key finding from multi-institutional studies is that the optimal radiomic features differ between the peripheral zone and transition zone. Cancer in the peripheral zone involves different tissue characteristics than cancer in the transition zone, where benign prostatic hyperplasia creates confounding background changes. This means that zone-specific models may be necessary for optimal performance.
The MAPS model (morphology, asymmetry, physiology, size) by Cameron et al. combined high-level spatial features from tumor regions with mpMRI, achieving 87% accuracy, 86% sensitivity, and 88% specificity. This approach demonstrated that incorporating spatial and shape information alongside texture substantially improves diagnostic accuracy beyond texture features alone.
Distinguishing clinically significant cancer (Gleason Score 7 or above) from clinically insignificant cancer (Gleason Score 6) is the central decision in prostate cancer management, determining whether a patient needs treatment or can be safely monitored. This distinction cannot be made reliably from PSA or basic imaging alone.
Fehr et al. demonstrated that combining ADC map features and T2-weighted texture in a recursive feature elimination support vector machine distinguished Gleason 6 from Gleason 7 or above with 93% accuracy. Importantly, the same model distinguished Gleason 7 (3+4) from Gleason 7 (4+3), two subtypes with different prognosis, with 92% accuracy. A classifier using only mean ADC achieved only 58 to 59% accuracy, confirming the value of texture over scalar measurements alone.
Nketiah et al. showed in both initial and multicenter validation studies that T2-weighted texture features, specifically homogeneity and entropy, were significantly correlated with Gleason grade. These features capture whether the tumor appears uniformly dark and dense (aggressive) or more variable (lower grade), providing a biologically interpretable link between image appearance and tumor biology.
Cuocolo et al. found that the surface area-to-volume ratio (SAVR) derived from ADC maps predicted clinical significance with AUC of 0.78, outperforming other shape features. Higher-grade tumors tend to be more compact and spherical relative to their volume, reflecting the disorganized growth pattern of more aggressive disease.
PI-RADS 3 lesions represent one of the most difficult challenges in prostate cancer management. These are considered equivocal, meaning clinicians cannot determine from MRI alone whether significant cancer is present, leaving patients in a difficult position between watchful waiting and potentially unnecessary biopsy.
Giambelluca et al. developed a texture analysis model specifically targeting PI-RADS 3 lesions larger than 5 mm, achieving AUC of 0.77 on T2-weighted images and AUC of 0.81 on ADC maps for diagnosing clinically significant cancer within this ambiguous category. These results suggest radiomics can add meaningful discriminating power precisely where radiologist judgment provides the least certainty.
Brancato et al. extended this approach to both PI-RADS 3 and PI-RADS 4 lesions, achieving AUC of 0.80 and 0.89 respectively using texture features from T2-weighted and ADC images. A notable finding was that DCE-MRI features were not useful for building the predictive model, suggesting that texture patterns from structural and diffusion imaging contain more relevant prognostic information than vascular perfusion data for these borderline lesions.
These findings point toward a clinical workflow where radiomics scores augment rather than replace PI-RADS assessment, providing a quantitative second opinion on equivocal lesions. This could meaningfully reduce the number of PI-RADS 3 patients referred for biopsy while maintaining a safety margin for catching significant cancers.
Beyond diagnosis, radiomics has been applied to guiding radiation therapy and predicting treatment outcomes. Shiradkar et al. developed the Rad-TRaP framework, which uses radiomic feature analysis to precisely identify cancer locations on MRI, transfers these boundaries to CT for treatment planning, and generates targeted focal radiation plans. This approach could concentrate radiation on the tumor while reducing doses to healthy tissue.
Abdollahi et al. applied radiomics to monitor radiation-induced changes in the bladder wall during prostate cancer radiotherapy, finding that GLCM texture features changed measurably in proportion to radiation dose. This suggests radiomics could serve as a quantitative tool for measuring treatment toxicity and adjusting plans in response to tissue changes over the course of therapy.
Biochemical recurrence, detected by rising PSA after surgery or radiation, affects a significant proportion of high-risk patients and signals treatment failure. Gnep et al. demonstrated that geometrical tumor characteristics and Haralick texture features from T2-weighted images correlated with Gleason grade and were significantly associated with biochemical recurrence, suggesting that pre-treatment MRI features can predict which patients are at highest risk for relapse.
These treatment applications share a common principle: imaging features that encode tumor biology at the time of diagnosis contain information about how that tumor will behave during and after treatment. Extracting and quantifying these features with radiomics provides a non-invasive window into treatment prognosis.
Despite promising results, radiomics in prostate cancer faces significant obstacles to clinical adoption. The most critical is lack of reproducibility: different studies use different imaging protocols, segmentation methods, feature extraction software, and statistical approaches, making it difficult to compare results or validate findings across institutions.
Most existing studies are retrospective with small sample sizes, typically under 200 patients. When the number of extracted features far exceeds the number of patients, there is a high risk of overfitting, where models perform well in the training dataset but fail to generalize. Large prospective multicenter trials are needed to establish which radiomic features are truly reproducible and biologically meaningful.
Future directions include radiogenomics, which integrates radiomic features with genomic and proteomic tumor data. The underlying premise is that imaging features visible on MRI reflect molecular and genetic characteristics of the tumor, and combining these information streams could substantially improve prediction of aggressiveness, treatment response, and prognosis beyond what either approach achieves alone.
A multimodal approach combining MRI with PET/CT radiomics represents another emerging direction, particularly with PSMA-targeted tracers that are highly specific for prostate cancer. Combining functional molecular imaging with structural texture analysis could provide a more comprehensive picture of tumor biology, improving both initial staging and monitoring of disease recurrence.