On September 2, 2025, at 5:00 p.m., a multidisciplinary team led by Professor Shancheng Ren of the Second Affiliated Hospital of Naval Medical University (Shanghai Changzheng Hospital), in collaboration with Professor Shudong Zhang (Peking University Third Hospital), Professor Jie Li (The First Affiliated Hospital of Nanjing Medical University), Professor Liang Wang (Beijing Friendship Hospital), Professor Pei Nie (Qingdao University Affiliated Hospital), and Professor Lizhi Shao (Anhui University), published a high-impact study in Nature Cancer (IF = 28.5) titled 'An MRI-Pathology Foundation Model for Non-Invasive Diagnosis and Grading of Prostate Cancer.'Research Background
Prostate cancer is the second most common malignancy among men worldwide. In China, its incidence has been rising steadily due to an aging population and increasingly Westernized lifestyles, and it now ranks sixth among male malignancies. Clinical data show that among men over 50 undergoing routine health examinations, one-third are found to have prostate nodules on ultrasound, and nearly 10% have elevated PSA levels — a finding that causes significant psychological distress for patients. Current global clinical guidelines recommend magnetic resonance imaging (MRI) with PI-RADS scoring for further evaluation. However, PI-RADS scoring faces two major limitations: 1. Subjectivity – The scoring depends heavily on radiologists’ experience, with inter-observer variability as high as 30%. 2. Limited accuracy – Even when an experienced radiologist provides a PI-RADS score, it cannot reliably determine the presence or absence of malignancy. These limitations highlight the urgent need for a highly efficient, accurate, and non-invasive diagnostic and grading tool to assist in evaluating patients with suspected prostate cancer.
Research Findings
The team developed and validated a multi-center, image-pathology foundation model for the efficient, accurate, and non-invasive diagnosis and grading of prostate cancer. This work demonstrated how artificial intelligence (AI) combined with MRI can quantitatively capture pathological features of prostate tumors. The model has the potential to significantly reduce the need for invasive prostate biopsies, lower procedure-related pain and complications, and ultimately improve the overall diagnostic experience for patients.
A total of 5,747 patients were recruited from multiple centers, including both retrospective and prospective cohorts. Radiological, pathological, and clinical data were collected, and the AI model was evaluated using temporal, spatial, and population-based external validation as well as prospective testing. To address challenges such as missing sequences, overfitting, and inter-scanner variability, the team used 1,296,950 paired imaging samples to train the foundation model, integrating self-supervised learning, multi-task learning, Transformer architecture, and transfer learning techniques to markedly enhance predictive performance.
The AI model — named MRI-Based Predicted Transformer for Prostate Cancer (MRI-PTPCa) — uses three MRI sequences (T2WI, DWI, ADC) to predict tumor aggressiveness, a parameter typically only obtainable through pathological assessment. This enables clinicians to diagnose prostate cancer, clinically significant prostate cancer, and perform pathological grading non-invasively. In the retrospective phase, seven institutions, four medical centers, and one international dataset participated in multi-center testing. In the prospective phase, MRI-PTPCa was tested as an independent system, a parallel decision support tool, and an early-warning system. The model demonstrated excellent concordance with pathological evaluation (P < 0.001) and outperformed clinical assessments and other predictive models (AUC for prostate cancer = 0.983, 95% CI 0.980–0.986; AUC for clinically significant prostate cancer = 0.978, 95% CI 0.975–0.980; grading accuracy = 89.1%, 95% CI 88.2–89.9%). Notably, when combined with multiparametric MRI, MRI-PTPCa’s performance for non-invasive diagnosis and grading was nearly equivalent to that of pathological evaluation.
Model Interpretability and Biological Relevance
To ensure clinical interpretability, the researchers compared MRI-PTPCa outputs with prostatectomy whole-mount pathology, AI-generated heatmaps, and quantitative imaging features. The results revealed a strong positive correlation between MRI-PTPCa scores and true Gleason grades. Attention heatmaps generated using Class Activation Mapping (CAM) highlighted the key regions contributing to predictions, while quantitative analyses confirmed the relevance of T2WI, DWI, and ADC sequences, showing strong alignment with expert PI-RADS consensus. Furthermore, the fused features within MRI-PTPCa were found to correlate significantly with tumor cellularity, morphology, and texture (p < 0.01), supporting the feasibility of linking imaging to pathology. MRI-PTPCa features were able to distinguish among non-cancerous, cancerous, and clinically significant prostate cancer phenotypes. The model’s encoded representations also correlated strongly with molecular markers such as tPSA, fPSA, and f/t PSA, offering additional molecular insights into prostate cancer biology.
Conclusion
This study represents a major step toward precision, non-invasive prostate cancer care. By quantitatively linking MRI features with pathological information, the MRI-PTPCa model enhances the ability of AI foundation models to deliver highly accurate, patient-friendly cancer diagnosis and grading in real-world clinical practice. Its implementation could substantially reduce unnecessary biopsies and improve safety and comfort for prostate cancer patients worldwide.
