Prostate cancer is a leading cause of cancer death in men globally, with over 1.4 million new cases and approximately 370,000 deaths per year according to GLOBOCAN 2022. Early detection is critically important: the 5-year survival rate for localized prostate cancer reaches 99%, while for metastatic disease it drops below 30%.
Three primary imaging modalities support prostate cancer diagnosis and management, each with distinct strengths and limitations. Transrectal Ultrasound (TRUS) is inexpensive and real-time but operator-dependent and poor at detecting small lesions. Multiparametric MRI (mp-MRI) is the gold standard for characterizing clinically significant cancer but is time-consuming and shows substantial reader variability across centers (Kappa 0.4-0.6). PSMA PET/CT excels at detecting metastatic lesions throughout the body but cannot identify small lesions under 5 mm and requires expensive equipment.
AI, including machine learning, deep learning, and radiomics, has the potential to enhance every step of imaging-based prostate cancer diagnosis: improving lesion detection in ultrasound, standardizing MRI interpretation, and automating the quantification of metastatic burden on PET/CT. This review systematically surveys AI applications across all three modalities while addressing clinical translation challenges and future directions.
Traditional gray-scale ultrasound has a specificity of only 55-65% for prostate cancer because cancer lesions (hypoechoic areas) overlap with benign conditions like prostatic hyperplasia and prostatitis. AI-based approaches address this by extracting features invisible to the naked eye. One computer-aided detection system combining multiresolution texture features with clinical descriptors achieved 92-96% sensitivity and 90-95% specificity on TRUS images.
Deep learning-based automatic lesion detection using YOLOv5 models can identify suspicious areas in real-time with a detection sensitivity of 92.3% and only 3.1 false positives per case, with individual image analysis in just 0.8 seconds -- far faster than radiologist assessment. A 3D convolutional neural network trained on whole-prostate TRUS videos (3D P-Net) achieved an AUC up to 0.90 with high accuracy across independent validation cohorts.
Radiomics-based models combining ultrasound features with PSA-related indicators have shown particular value in the PSA gray zone (4-10 ng/mL) -- the clinically ambiguous range where traditional methods have a misdiagnosis rate as high as 40%. One random forest model achieved an AUROC of 0.89 for distinguishing cancer from benign prostate hyperplasia, with the AUROC remaining 0.83 in the gray zone -- far superior to conventional ultrasound (0.72) or PSA alone (0.68).
Shear Wave Elastography measures tissue stiffness (cancer tissue is stiffer than normal tissue), and an AI model combining elastography with texture features achieved an AUROC of 0.91 with 94.5% detection rate for high-risk cancer (Gleason score 8 or above), providing a new quantitative risk stratification index.
The prostate has three anatomically distinct zones: the Peripheral Zone (PZ) where 70% of cancers arise, the Transition Zone (TZ) where 30% develop (and are harder to detect due to overlap with benign hyperplasia), and the Central Zone. Traditional manual MRI segmentation takes about 20 minutes per case with a Kappa consistency of only 0.5-0.7. Deep learning models based on U-Net++ architecture reduce this to 1.2 minutes with significantly improved accuracy, with TZ segmentation DSC reaching 0.90 after adding attention gating.
For identifying clinically significant prostate cancer (csPCa), a Vision Transformer-based multi-sequence fusion model (ViT-MRI) integrating T2-weighted, diffusion-weighted, and dynamic contrast-enhanced sequences achieved an AUROC of 0.94 with 91.2% sensitivity and 88.5% specificity on a 1,500-patient dataset -- significantly outperforming three senior radiologists (AUROC 0.85-0.88). For small cancers under 1 cm -- the most commonly missed -- AI detection rate was 87.3% versus radiologists' average of 72.1%.
Dynamic contrast-enhanced MRI (DCE-MRI) provides information about tumor vascularity and blood flow. AI models using long short-term memory (LSTM) networks can automatically classify time-intensity curve patterns and predict pharmacokinetic parameters (Ktrans, kep). The Ktrans value predicted by AI was positively correlated with Gleason score, and patients with AI-predicted Ktrans above 0.8 min^-1 had a 3.2-fold higher risk of biochemical recurrence.
PSMA PET/CT uses radiolabeled tracers targeting prostate-specific membrane antigen (PSMA), which is overexpressed in prostate cancer cells, to detect metastatic lesions throughout the body with high sensitivity. However, traditional image reading relies on subjective assessment of metabolically active areas, and small lesions under 5 mm are frequently missed.
A 3D-UNet model trained on 800 cases of 68Ga-PSMA-11 PET/CT (including 320 metastatic and 480 non-metastatic patients) achieved sensitivity and specificity of 90.5% and 92.3% for bone metastases and 88.1% and 91.7% for lymph node metastases -- outperforming two nuclear medicine physicians. The model automatically calculated standardized uptake values (SUVmax) and metabolic tumor volume (MTV) for each detected lesion.
MTV calculated by AI showed a significant inverse correlation with progression-free survival (hazard ratio 1.85), establishing it as a valuable quantitative biomarker for treatment planning. For monitoring treatment response, a contrastive learning model based on Siamese networks predicted 3-month treatment response from 1-month PET/CT data with an AUROC of 0.90, compared to only 0.75 for the traditional approach of comparing SUVmax change before and after treatment.
Three interconnected challenges limit the transition from research to clinical practice. The first is data quality and sharing: most AI models were trained on data from large academic centers, where imaging protocols, patient demographics, and reporting standards may differ substantially from community hospitals and emerging healthcare systems. Models trained on high-quality academic datasets often degrade significantly in real-world deployment settings.
Model generalization across different scanner vendors, field strengths (1.5T vs 3T), and acquisition protocols remains a fundamental technical challenge. A model achieving excellent performance on Siemens scanners may perform poorly on GE or Philips equipment. Addressing this requires either diverse multi-vendor training datasets or domain adaptation techniques that can automatically adjust to new acquisition environments.
Clinical integration involves both technical and human factors. AI tools must fit seamlessly into existing workflows -- connecting to Picture Archiving and Communication Systems (PACS), operating at acceptable speeds, and presenting results in formats that radiologists find useful rather than disruptive. Beyond the technical interface, building physician trust requires explainable AI outputs that help clinicians understand why the model made a particular recommendation, rather than simply displaying a probability score.
Ethical and regulatory considerations include ensuring algorithmic fairness (AI must perform equitably across racial, ethnic, and age groups), data governance (patient privacy protections for the large training datasets required), and liability frameworks (when AI contributes to a diagnostic error, who bears responsibility -- the radiologist, the institution, or the AI vendor). These governance questions must be resolved in parallel with technical development for safe clinical deployment.
The most transformative near-term opportunity is the integration of imaging AI with multi-omics data -- combining image-derived features with genomic, transcriptomic, and proteomic profiles. This multimodal approach could enable AI systems to predict not just whether cancer is present, but its specific molecular subtype, likely treatment sensitivity, and metastatic potential -- moving from diagnostic to predictive precision medicine.
Explainable AI (XAI) methods are essential to move beyond black-box predictions. Gradient-based visualization (showing which image regions most influenced the AI decision), attention maps, and counterfactual explanations (what would need to change for the AI to give a different answer) are being developed to make AI reasoning transparent. For prostate cancer imaging, XAI should identify specific PI-RADS-relevant features -- irregular margins, restricted diffusion, early enhancement -- that drove the model's risk assessment.
The vision for mature AI integration is not a standalone tool consulted separately, but a workflow-embedded decision support system that automatically pre-processes images, highlights suspicious regions, generates risk scores, and presents patient-specific context (prior PSA trend, biopsy history, genetic risk factors) -- all within the radiologist's normal reporting interface, at the speed of clinical practice.