Prostate cancer (PCa) is the most common cancer and the second leading cause of cancer-related death in men worldwide. Accurate identification of clinically significant prostate cancer (csPCa) -- defined as Gleason grade group 2 or higher -- before biopsy is critical to ensure timely treatment while avoiding unnecessary procedures for patients with lower-risk disease.
Multiparametric MRI (mp-MRI) is the current guideline-endorsed imaging method for optimizing biopsy decisions, offering high accuracy through the PI-RADS scoring system. However, its adoption remains limited by high cost (approximately 15 times the cost of ultrasound equipment), limited scanner availability, and the need for specialist expertise to interpret the images.
Transrectal ultrasound (TRUS) has been used for prostate examination for over 40 years. It is far cheaper, portable, and widely available -- but its traditional diagnostic performance is poor, with sensitivities ranging from 8% to 88% and wide variability in specificity. Its interpretation suffers from operator dependence and low reproducibility, limiting its clinical utility for cancer detection.
The research question this study addresses is whether a deep learning model trained on standardized TRUS video of the entire prostate can close this performance gap -- making prostate cancer detection as accurate as mp-MRI but using a more accessible and affordable imaging modality.
This was a prospective study enrolling 832 patients across four hospitals in China who underwent prostate biopsy between January 2021 and December 2022. All patients had a standardized end-to-end grayscale TRUS video of the complete prostate recorded before biopsy -- a continuous transverse scan from the base to the apex capturing the entire gland.
The training cohort (559 patients) came from Shanghai Tenth People's Hospital, with an internal validation cohort (140 patients) from the same institution. The external validation cohort (133 patients) came from three independent hospitals -- providing a rigorous test of how well models generalize across different ultrasound equipment manufacturers, operators, and patient populations.
Two AI models were developed: a 2D P-Net that analyzes individual video frames independently and aggregates their predictions, and a 3D P-Net that processes the entire video simultaneously to learn both spatial features within each frame and temporal relationships between consecutive frames. Both used ResNet50 as the backbone architecture.
The ground truth for all predictions was pathological confirmation of cancer grade from biopsy tissue or surgical specimens. Performance was measured using area under the receiver operating characteristic curve (AUC), required biopsy rate, and unnecessary biopsy rate -- a clinically meaningful metric that counts how many patients without csPCa were still sent to biopsy unnecessarily.
Before the CNN models could be trained, an automated prostate segmentation step was needed to isolate the gland from surrounding tissues in each video frame. A custom 2D segmentation network called Efficient-UNet was developed for this purpose, achieving a Dice coefficient of 0.91 on validation images -- comparable to manual expert segmentation.
Both P-Net models incorporated three specialized modules beyond the base ResNet50 backbone. The Feature Pyramid Network (FPN) enables the model to simultaneously use fine spatial details from early network layers and high-level semantic understanding from deeper layers, helping it handle the variable sizes of prostates across patients.
The Squeeze-and-Excitation Network (SEN) adaptively reweights the importance of each feature channel and each video frame, helping the model focus on the most informative spatial locations and temporal moments in the TRUS video. Self-supervised pretraining was also incorporated to improve performance despite the relatively limited labeled data.
The key architectural distinction between 2D and 3D P-Net is how video information is processed. 2D P-Net analyzes frames independently and aggregates their outputs through a decision algorithm, while 3D P-Net uses three-dimensional convolutional filters that jointly learn spatial features and temporal relationships across the entire video sequence -- more closely resembling how a radiologist reviews a complete volumetric scan.
In the internal validation cohort, 3D P-Net achieved an AUC of 0.89 (95%CI: 0.83, 0.95) for identifying csPCa. In the external validation cohort, it achieved an AUC of 0.85 (95%CI: 0.78, 0.93). These results were statistically similar to experienced radiologists using mp-MRI PI-RADS v2.1 (AUC: 0.83-0.86), confirming that AI-enhanced TRUS can match MRI-level diagnostic accuracy.
In contrast, the traditional TRUS 5-point Likert score system scored by experienced radiologists achieved AUCs of only 0.78 and 0.78 in the internal and external cohorts. 3D P-Net significantly outperformed both junior and experienced radiologists on TRUS scoring (p = 0.003-0.040) -- demonstrating that AI resolves the fundamental limitation of human TRUS interpretation.
2D P-Net showed strong internal validation performance (AUC 0.86) but larger performance variation across external cohorts (AUC range 0.73-0.92). 3D P-Net was more consistent across institutions (AUC range 0.84-0.91), confirming that learning temporal-spatial relationships from the full video produces a more robust and generalizable model than analyzing individual frames independently.
The models also showed clinically sensible PSA-stratified behavior: positive predictive value (PPV) increased in patients with high PSA levels (higher cancer prevalence) and negative predictive value (NPV) increased in patients with low PSA levels -- suggesting the AI captures biologically meaningful patterns rather than just statistical correlations.
The ability to reduce unnecessary biopsies is a primary clinical goal of AI-assisted prostate imaging. When using traditional TRUS 5-point Likert scoring, 40.3% of patients required biopsy with 38.1% of those biopsies proving unnecessary (benign or insignificant cancer). Using mp-MRI PI-RADS, 47.6% required biopsy with 35.2% unnecessary.
3D P-Net reduced the required biopsy rate to 34.0% while cutting unnecessary biopsies to just 25.8% -- substantially better than both traditional approaches. 2D P-Net also reduced biopsy rates but had a higher rate of missed csPCa diagnoses (17.6%), making it less clinically safe than 3D P-Net.
Decision curve analysis (DCA), which measures the net clinical benefit of using each strategy across a range of risk thresholds, confirmed that 3D P-Net provided the highest net benefit of all compared methods. This means that across different physician thresholds for deciding when to recommend biopsy, 3D P-Net consistently guided better decisions.
The models were also able to correctly classify as non-csPCa 87.1% of patients who had been flagged as suspicious by mp-MRI PI-RADS v2.1 -- suggesting 3D P-Net could serve as a confirmatory second read to reduce unnecessary biopsies even after MRI evaluation, not just as an alternative to MRI.
A key concern with clinical AI adoption is whether clinicians can understand and verify how the model is making decisions. To address this, the researchers used Grad-CAM (gradient-weighted class activation mapping) to generate heatmaps highlighting which regions of each TRUS frame contributed most to the csPCa prediction.
Visual inspection of the heatmaps showed that both 2D and 3D P-Net models focused on regions that roughly corresponded to the locations of pathologically confirmed csPCa in the prostate -- confirming that the models were detecting biologically relevant features rather than imaging artifacts or incidental patterns.
The agreement between the P-Net heatmap regions and the suspicious lesion locations identified by prostate mp-MRI was quantified with a Dice coefficient of 0.54. Given the inherent differences in imaging direction and resolution between TRUS and MRI, this level of spatial agreement is considered acceptable and supports the biological interpretability of the AI's decisions.
This heatmap capability has direct clinical utility: if available before biopsy, it could theoretically guide urologists to target the most suspicious regions identified by the AI model -- potentially improving biopsy yield and reducing sampling errors in the largest prostate glands where comprehensive coverage is most challenging.
The most important finding of this study is that a deep learning model applied to low-cost, widely available grayscale TRUS video can achieve diagnostic accuracy matching expensive mp-MRI interpreted by experienced radiologists. This has direct implications for healthcare equity in prostate cancer detection, particularly in settings where MRI is not accessible.
The prospective, multi-institutional study design is a critical strength, as it mirrors real-world clinical conditions much more closely than retrospective single-center studies. The fact that 3D P-Net maintained consistent performance across four hospitals using different TRUS equipment manufacturers demonstrates practical generalizability.
Adding clinical parameters such as PSA level and free PSA to the AI models did not significantly improve performance beyond imaging alone (AUCs changed minimally), suggesting that the deep learning models are already capturing the relevant biological information available from PSA values through the imaging data itself.
Important limitations include the absence of contrast-enhanced or multiparameter ultrasound (which could further improve sensitivity), the prospective but non-randomized design, and the need for further study on how the model performs in ethnic groups beyond the Chinese patients enrolled. Randomized trials comparing patient outcomes under AI-guided versus standard biopsy strategies are needed before widespread adoption.