Prostate cancer diagnosis typically requires a needle biopsy, which carries risks including infection, bleeding, and pain. However, because MRI readings and PSA levels are imperfect predictors of cancer presence, many men undergo biopsy that ultimately reveals no clinically significant cancer -- exposing them to unnecessary risk and discomfort.
Clinically significant prostate cancer (csPCa) is defined as Gleason Grade Group 2 or higher (Gleason score 3+4 or greater) -- cancers that pose a meaningful risk of progression and mortality. Clinically insignificant cancers (Grade Group 1) are typically managed with active surveillance rather than treatment. A key clinical challenge is accurately distinguishing between these two categories before biopsy.
The PI-RADS (Prostate Imaging-Reporting and Data System) scoring system provides radiologists with a standardized scale from 1 to 5 for reporting how suspicious a prostate lesion appears on MRI. However, PI-RADS scoring has significant inter-reader variability, and a substantial proportion of PI-RADS 3 lesions (indeterminate category) are clinically insignificant, making biopsy decisions in this group particularly difficult.
This study developed the ClaD nomogram -- an integrated clinical decision tool combining deep learning analysis of biparametric MRI with PI-RADS scoring and clinical laboratory values -- to more accurately stratify patients and reduce unnecessary biopsies.
The study enrolled 592 patients with 823 MRI-suspicious lesions from five institutions across the United States and the Netherlands. Patients underwent biparametric MRI (bpMRI) -- combining T2-weighted imaging and ADC maps derived from diffusion-weighted imaging -- without intravenous contrast injection. This simplified protocol is faster, cheaper, and avoids gadolinium contrast exposure.
The dataset was split by institution: three sites formed the training and cross-validation set, and two independent sites formed the external validation set. This geographic and institutional separation provides stronger evidence for generalizability than splitting a single-site dataset into train/test subsets.
Ground truth was determined by histopathology from MRI-targeted biopsy, systematic biopsy, or radical prostatectomy specimens. To maximize safety, the clinical threshold for significance was Gleason Grade Group 2 (GGG2) or higher -- meaning cancers where active treatment is typically recommended rather than surveillance.
Three clinical variables available before biopsy were incorporated into the nomogram: PSA (prostate-specific antigen), prostate volume (used to calculate PSA density), and lesion volume estimated from the MRI. These variables were chosen because they are routinely collected at most prostate cancer diagnostic centers and add complementary information to imaging features.
The deep learning imaging predictor (DIN) used convolutional neural network architectures including AlexNet and DenseNet, pre-trained on natural images and fine-tuned on prostate MRI. The network received three input channels: T2-weighted images, ADC maps, and binary lesion segmentation masks identifying the lesion boundaries.
A key innovation was extracting features not just from inside the lesion but also from the peritumoural region -- the 3 mm tissue margin surrounding the lesion boundary. The model was evaluated at multiple scales (lesion-only, lesion plus 1 mm, 3 mm, and 5 mm margins), and the 3 mm peritumoural scale proved most informative, suggesting that tissue changes at the tumor margin carry important diagnostic signals.
Model interpretability was assessed using Grad-CAM (Gradient-weighted Class Activation Mapping), a technique that highlights which image regions most influenced each prediction. Grad-CAM maps confirmed that the deep learning model focused on biologically meaningful regions -- the lesion core and immediately adjacent tissue -- rather than irrelevant background structures.
The final ClaD nomogram was constructed by logistic regression combining: (1) the deep learning imaging prediction score, (2) the PI-RADS score assigned by a radiologist, (3) PSA, (4) prostate volume, and (5) lesion volume. This ensemble approach was designed to leverage the complementary strengths of automated image analysis, expert clinical judgment, and serum biomarkers.
On the external validation cohort, the ClaD nomogram achieved an AUC of 0.81, significantly outperforming the deep learning imaging predictor alone (DIN, AUC 0.74) and PI-RADS-based nomogram alone (PIN, AUC 0.76), with both differences reaching statistical significance (p less than 0.001). The full nomogram outperforms its component parts.
At a decision threshold preserving 90% sensitivity for csPCa, ClaD could avoid 59.18% of unnecessary biopsies in the validation cohort. Patients classified by ClaD as very low, low, and intermediate favorable risk had progressively decreasing rates of csPCa on biopsy, confirming that the nomogram meaningfully risk-stratifies patients across the clinical spectrum.
Adding PI-RADS to the deep learning model increased AUC by 7% compared to DIN alone. This confirms that expert radiologist interpretation of MRI retains independent diagnostic value beyond what the automated model can extract, supporting a collaborative human-AI rather than full automation approach for clinical deployment.
In a subset of 81 patients who underwent radical prostatectomy, the ClaD nomogram predictions correlated with biochemical recurrence-free survival (AUC 0.83 for prostatectomy specimens). This secondary finding suggests the nomogram captures not just cancer presence but also cancer aggressiveness -- an important bonus for long-term patient management decisions.
The study's most important methodological strength is its multi-institutional design with geographically separate validation sites. Most prostate AI studies validate on internal cohorts, which overestimates real-world performance. Testing on patients from two additional institutions confirms that ClaD generalizes across different scanner hardware, radiologist practices, and patient populations.
The finding that peritumoural regions contribute significant discriminative information aligns with known cancer biology: invasive cancers remodel the extracellular matrix and induce changes in surrounding tissue that are detectable on MRI. Restricting analysis to the lesion core alone misses this important contextual signal.
A limitation is that the study relied on manual prostate gland and lesion delineation by radiologists at each institution. This is time-consuming and introduces inter-reader variability in lesion volume estimates, which could affect nomogram predictions. Automated lesion detection and segmentation would be needed for practical clinical deployment.
The inclusion of PI-RADS scoring in the nomogram means the system requires both AI analysis and expert radiologist reading -- it does not replace the radiologist but augments their assessment. This human-AI collaboration model may actually be more acceptable to clinicians and regulators than fully autonomous systems, even if it is slightly less efficient.
The ClaD integrated nomogram provides a non-invasive, pre-biopsy risk stratification tool that could reduce unnecessary biopsies in a substantial proportion of patients while maintaining near-complete sensitivity for clinically significant cancer. This addresses a major unmet need in prostate cancer diagnosis.
Beyond diagnosis, the nomogram's association with biochemical recurrence-free survival opens the possibility of using ClaD to identify patients at high risk of post-treatment relapse who might benefit from adjuvant therapy after definitive treatment. This dual-use -- triage for biopsy and prognosis after treatment -- maximizes the clinical value of a single MRI-based assessment.
Future work should incorporate automatic segmentation pipelines to eliminate manual contouring, extend validation to include digital rectal examination findings and free-to-total PSA ratios that were unavailable across all participating sites, and test whether ClaD performance holds in ethnically diverse patient populations.
Wider validation across more global institutions and prospective clinical trials testing whether ClaD-guided biopsy decisions improve patient outcomes compared to standard PI-RADS-based clinical practice will be necessary before regulatory approval and routine clinical adoption.