Biochemical recurrence (BCR) -- a rising PSA level after radical prostatectomy -- affects 20 to 40% of patients treated for prostate cancer and is the earliest signal that the cancer may be coming back. Early identification of high-risk patients allows timely addition of radiation or hormone therapy, potentially preventing metastatic spread.
Current risk stratification tools include CAPRA-S (a clinical score incorporating PSA, Gleason grade, and pathological stage), NCCN guidelines, and the Gleason Grade Group from the radical prostatectomy specimen (GG-RP). These tools use post-operative pathology, but they have limited accuracy -- their concordance indices for predicting BCR range from roughly 0.58 to 0.68, leaving substantial room for improvement.
MRI-derived radiomic features -- quantitative measurements extracted from prostate MRI scans taken before surgery -- offer a potentially powerful complement to pathological data. Radiomics can detect subtle tissue characteristics invisible to the naked eye, such as heterogeneity in tissue texture that correlates with tumor aggressiveness.
This multi-center study from three Beijing hospitals developed and validated a deep learning radiomic model (DRS-BCR) for predicting biochemical recurrence-free survival, combining radiomic feature extraction with deep survival neural network architecture.
The study enrolled 485 patients from three Beijing institutions who underwent radical prostatectomy and had preoperative T2-weighted MRI (T2WI). One institution provided the primary training cohort (296 patients); the other two institutions provided independent external validation cohorts (96 and 93 patients).
Radiologists manually delineated the prostate on each MRI scan, and 702 radiomic features were extracted from these regions of interest using a standardized radiomic pipeline. Features captured first-order statistics (intensity distribution), texture measures (GLCM, GLRLM patterns), and shape characteristics of the segmented prostate volume.
To reduce dimensionality and avoid overfitting, least absolute shrinkage and selection operator (LASSO) regression was applied to select the most prognostically relevant features. This penalized regression method shrinks uninformative feature coefficients to zero, leaving only features with the strongest associations with BCR outcomes.
Two types of annotation regions were tested: a dense box (rectangular bounding box around the prostate) and an auto-coding box (a compressed representation learned by an autoencoder). The deep learning model incorporated both to capture both local and global textural information.
The core model is a Deep Radiomic Survival-BCR (DRS-BCR) network combining radiomic features with a deep survival neural network (DSNN) architecture. DSNNs extend traditional survival analysis (Cox proportional hazards) by learning complex, non-linear relationships between features and time-to-event outcomes -- in this case, time to biochemical recurrence.
The network learned a risk score for each patient that predicts their probability of remaining recurrence-free over time. This score is continuous rather than categorical, allowing fine-grained risk stratification rather than just sorting patients into low/intermediate/high risk buckets.
In addition to radiomic features, the model incorporated preoperative clinical variables including PSA level, Gleason grade from biopsy, and clinical T-stage. A separate model variant (DRC-BCR) also included postoperative pathological variables such as surgical margin status, though the preoperative model is more clinically valuable because it can inform decision-making before surgery.
The model was trained using the Adam optimizer with early stopping to prevent overfitting, and performance was evaluated using the concordance index (C-index) -- a measure of how well the model's risk scores rank patients by actual recurrence-free survival time, where 1.0 is perfect and 0.5 is random chance.
In the primary training and internal validation cohort, DRS-BCR achieved a C-index of 0.802 -- substantially better than the clinical comparators: CAPRA-S (0.677), NCCN (0.586), and GG-RP (0.583). This performance advantage was maintained in both external validation cohorts from the other two Beijing institutions.
Kaplan-Meier survival analysis confirmed that patients stratified into high and low risk groups by DRS-BCR had significantly different recurrence-free survival curves (p less than 0.001). The separation between groups was wider and cleaner than when patients were stratified using CAPRA-S alone, demonstrating that the radiomic information adds meaningful independent prognostic value beyond clinical variables.
A head-to-head comparison against CAPRA-S showed that DRS-BCR identified approximately one-third fewer false positives -- patients predicted to recur who actually did not. This reduction in unnecessary post-surgical treatment (such as adjuvant radiation) has direct clinical implications for patient quality of life and healthcare resource utilization.
The postoperative model (DRC-BCR) including surgical pathology variables performed marginally better (C-index 0.816), confirming that adding post-surgical information incrementally improves prediction, but the key finding is that the preoperative DRS-BCR is already highly informative for treatment planning before the patient undergoes surgery.
The study's multi-center design -- with three independent institutions -- is a major strength. AI models validated only at a single site often fail to generalize because they overfit to the specific scanner, acquisition protocol, or patient demographics of one institution. The consistent C-index values across all three Beijing cohorts suggest the model captures generalizable tumor biology rather than site-specific imaging artifacts.
The ability to predict BCR from preoperative MRI has important implications for surgical planning. Patients identified as high-risk before surgery could be counseled about the likelihood of needing adjuvant treatment, allowing earlier enrollment in clinical trials of neoadjuvant therapies and potentially improving outcomes.
A limitation is that all three institutions are academic medical centers in Beijing, and the patient population (predominantly East Asian) may not represent the full global diversity of prostate cancer biology. The model's performance in Western populations or in patients with different ethnic backgrounds remains to be validated.
The study used only T2-weighted MRI as imaging input. Incorporating additional sequences such as diffusion-weighted imaging (DWI) or dynamic contrast enhancement could potentially improve model performance further, as these sequences provide complementary information about tumor cellularity and vascularity.
The DRS-BCR model demonstrates that preoperative T2-weighted MRI contains prognostic information beyond what current clinical staging and pathological grading capture, and that deep learning can extract this information reliably across multiple institutions.
The practical clinical value is clear: identifying patients at high risk for BCR before surgery allows personalized treatment planning -- including decisions about nerve-sparing surgery technique, lymph node dissection extent, and eligibility for neoadjuvant clinical trials -- that is not currently supported by existing risk calculators.
Integration of this model into clinical workflows would require automatic prostate segmentation tools to replace the manual contouring used in this study, which is time-consuming and subject to inter-observer variability. Pairing DRS-BCR with automated segmentation would create a fully automated preoperative risk prediction pipeline.
Future work should validate the approach across more diverse international cohorts, incorporate multiparametric MRI sequences, and test whether the model adds value on top of recently developed genomic biomarker panels -- creating a combined imaging-genomic risk stratification tool for prostate cancer management.