Prostate cancer (PCa) is the most frequently diagnosed cancer in men in 112 of 185 countries worldwide. Accurate diagnosis before biopsy is the critical first step in appropriate management, yet this remains a challenging clinical problem because many non-cancerous prostate conditions can look similar to cancer on imaging.
The standard diagnostic pathway relies on elevated PSA (prostate-specific antigen) and abnormal digital rectal examination, followed by transrectal ultrasound-guided biopsy. However, this approach leads to unnecessary biopsies in patients with benign conditions, and still misses some cancers. Multiparametric MRI (mpMRI) before biopsy is now recommended by clinical guidelines, and the PI-RADS (Prostate Imaging Reporting and Data System) scoring framework is widely used to standardize how suspicious lesions are reported.
A key limitation of PI-RADS is that many benign conditions (BCs) -- including benign prostatic hyperplasia, prostatitis, and normal anatomical variants -- show overlapping appearances with prostate cancer on MRI. Even among expert radiologists, the sensitivity and specificity for cancer diagnosis vary enormously: published ranges span 73-100% and 8-100%, respectively, reflecting substantial interreader variability.
Deep learning and machine learning models have shown promise in reducing this variability. However, many existing models lack external validation or suffer from overfitting -- performing well on training data but poorly on new patients. Transfer learning (TL) offers a solution by starting with knowledge learned from related tasks and adapting it to the specific problem, enabling high performance even with limited training data.
The study enrolled 709 patients from two Chinese hospitals with pathologically confirmed prostate cancer or benign conditions between 2015 and 2021. Patients were divided into a training cohort (309 patients), an internal validation cohort (200 patients from the same institutions), and an external validation cohort (200 patients from a separate institution). All patients had prebiopsy MRI with biopsy results as the ground truth.
All MRI scans were performed on 3 Tesla MRI scanners (Philips Ingenia and Siemens TrioTim) without an endorectal coil, using standardized protocols. MRI images were interpreted by two board-certified radiologists with different levels of expertise -- one classified as 'expert' and one as 'basic' according to European Society of Urogenital Radiology consensus criteria -- and lesions were scored according to PI-RADS v2.1. Disagreements were resolved by consensus discussion.
For deep learning analysis, the whole prostate was used as the region of interest (ROI), delineated by three radiologists on axial T2-weighted images. Images were preprocessed to a standardized 224 x 224 x 3 pixel format required by the transfer learning model. This whole-gland approach differs from lesion-targeted methods and allows the model to consider the full prostate context.
The transfer learning model used in this study is built on a hybrid architecture combining a convolutional block (for extracting local image features) and an MLP Mixer layer (for integrating information across spatial regions and image channels). This design allows the model to capture both fine-grained local patterns and broader contextual relationships within the prostate MRI.
A key innovation is the choice of source domain for transfer learning. Rather than starting from a model trained on natural photographs (like ImageNet), the researchers pre-trained the model using whole slide images (WSIs) of prostate cancer from the Cancer Genome Atlas. These pathology slides contain detailed histological patterns of prostate cancer, which share underlying biological characteristics with MRI appearance. The researchers hypothesized -- and confirmed -- that transferring knowledge from cancer pathology slides to MRI would be more effective than transferring from unrelated natural images.
From each patient's prostate MRI, the model extracted 28,320 transfer learning features. These high-dimensional features were then processed through LASSO (Least Absolute Shrinkage and Selection Operator) regularization with 10-fold cross-validation to select the most informative subset. From 21,900 features showing statistically significant differences between cancer and benign cases, LASSO selected 73 features with non-zero coefficients for inclusion in the final transfer learning signature (TLS).
Three different classifiers -- LASSO, Support Vector Machine (SVM), and Extreme Learning Machine (ELM) -- were evaluated to identify which best captured the diagnostic signal in the transfer learning features. LASSO consistently outperformed SVM and ELM across all cohorts, achieving the highest AUC values in training, internal validation, and external validation, making it the basis for the final model.
A nomogram is a visual tool used in medicine that combines multiple risk factors to estimate an individual patient's probability of having a disease. The Transfer Learning Nomogram (TLN) was constructed by combining three independent predictors identified through multivariate logistic regression: the TLS score (derived from the 73 selected features), the patient's PI-RADS score, and age.
The TLS had an overwhelming predictive contribution -- its odds ratio of 120.780 (95% CI 34.648 to 421.028) indicates that a one-unit increase in TLS score was associated with an approximately 121-fold increase in odds of prostate cancer. PI-RADS score added independent contribution (OR 3.472), as did age (OR 1.088 per year). Notably, PSA and prostate volume -- while significant in the pure clinical model -- did not add independent value once the TLS and PI-RADS score were included.
The nomogram provides a practical visual interface: a clinician enters the patient's age, PI-RADS score, and TLS score, and reads off an estimated cancer probability from a continuous scale. The paper provides two concrete examples showing how the nomogram correctly predicted benign pathology (total score 52.5, risk below 0.1) and clinically significant cancer (total score 85.1, risk above 0.9) in two patients with identical PI-RADS category 3 lesions but different TLS scores -- demonstrating the added discriminative value of the AI features.
The TLN achieved AUC values of 0.9757 in training, 0.9255 in the internal validation cohort, and 0.8766 in the external validation cohort for distinguishing prostate cancer from benign conditions. In the internal validation cohort, it achieved 86.4% sensitivity, 88.4% specificity, and 87.5% accuracy. In the more challenging external validation, it achieved 72.7% sensitivity, 83.5% specificity, and 77.0% accuracy.
Across both validation cohorts, the TLN significantly outperformed both the standalone Transfer Learning Signature (TLS) and the pure Clinical Model (CM). The DeLong test and Integrated Discrimination Improvement (IDI) confirmed that these differences were statistically significant (p < 0.001 and p = 0.0263 for the two cohorts respectively, with IDI values of 0.0611 and 0.1272), validating that the combination of AI features with clinical factors adds genuine incremental value.
Decision curve analysis (DCA) -- which evaluates the net clinical benefit of a model across the full range of risk thresholds -- showed that the TLN provided greater net benefit than the clinical model across risk thresholds from 0.01 to 0.99. This means that at virtually any threshold a clinician might use to decide whether to recommend biopsy, the TLN-based decision rule is more beneficial than relying on clinical factors and PI-RADS alone.
The confusion matrix for all 709 patients across cohorts showed 336 true negative (benign correctly identified as benign) and 280 true positive (cancer correctly identified), with 39 false positives and 54 false negatives. The reasonably balanced error distribution suggests the model does not dramatically favor one type of error over the other, which is important for clinical applicability.
The superiority of the TLN over the clinical model confirms a key principle: imaging-derived AI features capture information about the prostate that cannot be inferred from PSA levels or PI-RADS scores alone. While clinical parameters like PSA are affected by benign conditions (for example, benign prostatic hyperplasia also elevates PSA), the TLS features reflect subtle image-level patterns more specifically associated with cancer.
The choice to pre-train on prostate cancer pathology slides rather than natural images was a deliberate and effective strategy. The more similar the source domain is to the target domain, the better the knowledge transfers. Prostate histopathology and prostate MRI share underlying tumor biology -- both reflect the same cellular and tissue abnormalities -- giving the model a more relevant starting point than generic image classifiers.
The study acknowledges several limitations. The sample size, while larger than many comparable studies, may still be insufficient for routine clinical deployment, and further multicenter validation is needed. The pathological ground truth relied primarily on biopsy rather than prostatectomy specimens in most cases, which means some cancers may have been missed due to sampling error. Additionally, the model distinguishes cancer from benign conditions but does not further differentiate clinically significant from clinically insignificant cancers -- a distinction that is critical for avoiding overtreatment of low-grade disease.
The Transfer Learning Nomogram (TLN) represents a promising noninvasive tool for differentiating prostate cancer from benign conditions before biopsy. By combining AI-derived image features with clinician-assigned PI-RADS scores and patient age, the model captures complementary information that no single input can provide alone.
A practical benefit of the nomogram format is its interpretability and ease of use in clinical settings. Unlike opaque neural network black boxes, a nomogram explicitly shows the contribution of each input and produces a probability estimate that radiologists and urologists can use alongside their clinical judgment -- supporting, rather than replacing, clinical decision-making.
The ultimate goal of this work -- and the broader field of AI-assisted prostate cancer diagnosis -- is to reduce the number of unnecessary biopsies while ensuring that clinically significant cancers are not missed. As next steps, the researchers plan to develop a model capable of distinguishing clinically significant prostate cancer (csPCa) from insignificant disease, which would provide the most actionable guidance for treatment decisions.