Prostate cancer screening using PSA (prostate-specific antigen) blood tests has been shown to reduce cancer mortality, but it comes with a major drawback: overdiagnosis rates of 27-56%. Many PSA-detected cancers would never have caused symptoms or death, yet are treated aggressively, leading to unnecessary surgery, radiation, and side effects including incontinence and erectile dysfunction.
Additionally, PSA misses a substantial portion of real cancers -- between 15% and 44% of biopsy-proven cancers occur in men with PSA levels below 4 ng/mL, the threshold typically used to trigger a biopsy. This means relying on PSA alone can both over-catch harmless cancers and under-catch dangerous ones.
Prostate MRI has emerged as a more specific screening tool. Major clinical trials (PROMIS and PRECISION) showed that MRI before biopsy reduced unnecessary biopsies by 25% while detecting more clinically significant cancers. As a result, MRI has been incorporated into prostate cancer diagnostic guidelines worldwide.
The idea of an image-guided prostate cancer screening program -- analogous to mammography screening for breast cancer -- is increasingly discussed. However, reading large volumes of prostate MRI scans is time-consuming and requires specialized expertise. Automating this process with AI could make image-guided screening feasible at scale.
This prospective proof-of-concept study enrolled 49 healthy, biopsy-naive men over 45 years old with no prior cancer history. Rather than the full multi-parametric MRI (mpMRI) protocol typically used in clinical practice, all participants underwent a streamlined biparametric MRI (bpMRI) protocol that combined only T2-weighted images and diffusion-weighted imaging (DWI), with a total scan time of approximately 9.5 minutes.
Biparametric MRI is faster and cheaper than full multi-parametric MRI (which also includes a contrast-enhancing injection). Research suggests bpMRI provides similar diagnostic accuracy in selected patients, making it more suitable for a population-wide screening program where cost and patient throughput are critical factors.
All scans were read independently by two experienced board-certified radiologists using the standardized PI-RADS v2 scoring system, which rates lesions from 1 (almost certainly benign) to 5 (highly suspicious for clinically significant cancer). All lesions scored PI-RADS 3 or higher were biopsied using MRI-ultrasound fusion targeted biopsy.
The same scans were also analyzed by a prototype AI software called ProstateAI (Siemens Healthineers), which automatically performed prostate segmentation, lesion detection, false-positive reduction, and PI-RADS category assignment -- all within an average of 14 seconds per case.
The ProstateAI system processes raw MRI data through a sequential pipeline of three deep neural networks. The first is a Candidate Localization Network -- a fully convolutional network that generates a probability heatmap across the entire prostate volume, with hotter (redder) colors indicating higher likelihood of a suspicious lesion at each location.
The second network is a Candidate Qualification Network (false-positive reduction network), which examines each detected candidate in a smaller sub-volume and determines whether it is a true lesion or a false alarm. Without this step, the sensitivity would be very high (97%) but the false-positive rate would be unacceptably high (84% of negative cases flagged).
The third network is a Classification Network that assigns each surviving lesion candidate to a PI-RADS category (3, 4, or 5), matching the scoring system used by radiologists. This allows the AI output to be directly integrated into clinical workflows that already use PI-RADS.
The AI was trained on 2,170 biparametric MRI examinations from eight different institutions, with expert radiologists providing pixel-level lesion boundary annotations and PI-RADS scores. This multicenter training design was specifically intended to capture the variability in scanner hardware, patient demographics, and imaging protocols encountered across different hospitals.
On a case level, the AI achieved 87% sensitivity -- correctly flagging 33 of 38 cases that radiologists identified as having a suspicious lesion. Specificity was 50%, meaning the algorithm incorrectly flagged 5 of the 10 truly negative cases. The kappa agreement score of 0.42 (moderate agreement) closely matched the known interobserver agreement between human radiologists for PI-RADS scoring.
Detection performance varied systematically by PI-RADS category, as expected. The AI detected 100% of PI-RADS 5 lesions (the most aggressive), 73% of PI-RADS 4 lesions, and 43% of PI-RADS 3 lesions. When only the highest-risk (index) lesion per patient was considered, sensitivity was 78%, 93%, and 100% for PI-RADS 3, 4, and 5 categories respectively.
Most importantly, all 6 patients with biopsy-confirmed prostate cancer were correctly detected by both the human radiologists and the AI system. Three of these 6 patients had PSA levels below 4 ng/mL -- the standard threshold for biopsy -- meaning PSA-only screening would have missed them entirely. The AI and MRI together identified them despite low PSA.
The low cancer yield in PI-RADS 3 lesions (just 1/28, or 3.6%) is notable. In diagnostic cohorts with clinical suspicion, PI-RADS 3 cancer yields typically range from 12-33%. The much lower yield in this healthy screening population reflects the difference in baseline cancer probability between screening-eligible and symptom-driven populations.
The most significant clinical finding is that MRI -- whether read by humans or the AI -- detected all biopsy-proven cancers, including cases that would have been missed by PSA screening. Three of six cancer patients had PSA below 4 ng/mL, the conventional threshold. This supports the view that MRI-based screening could serve as a more accurate early warning system than PSA testing, particularly for detecting clinically significant cancers in otherwise healthy men.
The low false-positive rate (0.875 per patient) with the false-positive reduction network enabled is a key practical achievement. Earlier AI systems for prostate lesion detection produced far more false positives, which would generate excessive unnecessary biopsies in a screening program -- the very problem that discredits PSA screening.
The 14-second automated analysis time is transformative for scalability. A single radiologist reviewing prostate MRIs spends 20-30 minutes per case. An AI system that pre-reads each scan and flags only the highest-suspicion findings could dramatically increase the throughput of a screening program and reduce the radiologist time required per case.
For the approximately 51% of this cohort's lesions that were PI-RADS 3 (equivocal), neither human radiologists nor AI achieve high reliability. The authors suggest this raises questions about whether PI-RADS 3 lesions should trigger immediate biopsy in a screening population with low baseline cancer risk. Larger studies are needed to develop appropriate management strategies for this ambiguous category in healthy screenees.
This proof-of-concept study demonstrates that a fully automated AI system can detect and PI-RADS-classify suspicious prostate lesions on biparametric MRI with sensitivity and interobserver agreement comparable to experienced human radiologists, achieving 100% detection of all biopsy-confirmed cancers in a healthy screening cohort.
The results support the feasibility of integrating AI-assisted biparametric MRI into opportunistic prostate cancer screening programs, analogous to AI-assisted mammography screening. The combination of short scan times (~9.5 minutes) and rapid automated analysis (14 seconds) makes population-level deployment technically plausible.
Limitations of the current work include the small sample size (48 evaluable cases), restriction to peripheral zone lesions only (transition zone performance remains untested), and the proof-of-concept nature of the design. The lower kappa score compared to diagnostic cohorts reflects the genuine difficulty of scoring equivocal PI-RADS 3 lesions in a healthy population with low cancer prevalence.
The authors call for larger prospective validation studies comparing AI predictions to biopsy outcomes across the full range of PI-RADS scores, and recommend investigating whether the AI's ability to detect visual patterns invisible to human eyes might particularly improve detection of transition zone cancers -- a known challenge for radiologists.