Multiparametric MRI of the prostate has become a standard tool for detecting and risk-stratifying clinically significant prostate cancer (csPCa). Using the PI-RADS scoring system (version 2.1), radiologists rate lesions on a scale of 1 to 5. Scores of 1 and 2 indicate very low probability of significant cancer, while 4 and 5 indicate high probability.
PI-RADS category 3 is the problematic middle ground - it is defined as equivocal, meaning radiologists cannot confidently determine whether significant cancer is present. In current clinical practice, most PI-RADS 3 patients are sent for biopsy to avoid missed diagnoses.
However, research shows that the majority of PI-RADS 3 lesions do not contain clinically significant cancer. This creates a substantial problem: a large number of patients undergo invasive, uncomfortable biopsies unnecessarily, with associated risks including pain, infection, and anxiety.
This study developed a new type of deep learning model - named AttenNet - specifically designed to re-stratify PI-RADS 3 patients, separating those who truly need biopsy from those who can safely avoid it. It is the first application of deep learning specifically focused on differentiating among PI-RADS 3 lesions.
The study enrolled patients suspected of prostate cancer from six medical centers in China, covering cases from January 2015 to December 2020. A total of 1,567 patients with MRI-guided biopsy were included, of whom 423 (21.1%) had PI-RADS category 3 lesions - the focus of this study.
The training and pretraining cohort consisted of centers 1 through 3: 1,144 patients with PI-RADS 1, 2, 4, and 5 were used for pretraining, and 238 PI-RADS 3 patients were used for retraining. The external testing cohort consisted of 185 PI-RADS 3 patients from the other three centers (4, 5, and 6), which the model had never seen during development.
All patients underwent biparametric MRI at 3.0 Tesla, providing three imaging sequences: T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) maps. Notably, gadolinium contrast-enhanced MRI was not used, avoiding the risks and costs associated with contrast agents.
Histopathological confirmation was obtained from MRI-targeted biopsy combined with systematic transrectal ultrasound-guided biopsy. Clinically significant PCa was defined as ISUP grade group 3 or higher, a threshold reflecting tumors with meaningful malignant potential requiring treatment.
The AttenNet architecture consists of three parallel branches - one each for T2WI, DWI, and ADC images - each using a 3D ResNet as its base. Three-dimensional processing is important because prostate lesions exist in three-dimensional space, and 3D models can capture spatial relationships that 2D slice-by-slice approaches miss.
Two special attention mechanisms were added to make AttenNet sensitive to the subtle features of PI-RADS 3 lesions. The channel attention module automatically highlights the most discriminative image features while filtering out noise throughout training. The soft attention module incorporates the anatomical location of the lesion - peripheral zone (PZ) or transitional zone (TZ) - as clinical prior knowledge.
The anatomical location matters because PI-RADS 2.1 guidelines specify that T2WI and DWI scores are the dominant factors for lesions in the transitional zone and peripheral zone, respectively. By incorporating this clinical knowledge directly into the model architecture, AttenNet mirrors how experienced radiologists actually evaluate lesions.
A transfer learning strategy was used to address the challenge of limited PI-RADS 3 training data. The model was first pretrained on the easier-to-classify PI-RADS 1, 2, 4, and 5 cases to learn general prostate cancer features, then specifically retrained on PI-RADS 3 cases to fine-tune its ability to detect subtle differences in this equivocal category.
For predicting any prostate cancer (PCa) in PI-RADS 3 patients, the AttenNet model achieved AUC values of 0.795, 0.963, and 0.922 across the three external testing centers (centers 4, 5, and 6 respectively). At centers 4 and 5, these values were significantly higher than simpler models lacking the attention modules (p values all below 0.05).
For predicting clinically significant PCa specifically, AttenNet achieved AUCs of 0.827 (center 4) and 0.926 (center 5). At center 6, all PI-RADS 3 patients were confirmed as non-csPCa by pathology, so no csPCa AUC could be computed for that cohort. At center 5, the AttenNet outperformed the ResNet-based models without attention modules (p = 0.017 and p = 0.037).
Ablation experiments confirmed each component's value: adding the transfer learning module improved performance, adding channel attention added further improvement, and incorporating the soft attention with location prior achieved the best results. Each module contributed incrementally to the final AttenNet architecture.
Decision curve analysis confirmed that using the AttenNet model's biopsy strategy (biopsy only those the model flags as high-risk) provided greater net benefit to patients than the current practice of biopsying all PI-RADS 3 patients across all testing cohorts.
The most clinically important finding was how effectively AttenNet identified patients who did NOT have clinically significant cancer - those who could safely avoid biopsy. The model identified 71.1% to 92.2% of non-csPCa patients across the three testing centers, all of whom had previously been sent for biopsy under standard PI-RADS 3 protocols.
In practice, AttenNet re-classifies PI-RADS 3 patients into two groups: PI-RADS 3U (upgraded) - those with higher risk who should proceed to biopsy, and PI-RADS 3D (downgraded) - those with lower risk who can potentially be safely monitored with repeat MRI or PSA testing instead of immediate biopsy.
The current standard of practice biopsied 84.7% and 79.7% of PI-RADS 3 patients unnecessarily at the two primary testing centers. AttenNet's model-guided strategy could dramatically reduce this over-biopsy rate, lowering the proportion of unnecessary invasive procedures while maintaining high sensitivity for catching true cancers.
Subgroup analyses showed the model performed well across different tumor sizes and PSA levels, with satisfactory AUCs in most subgroups. Notably, the model was robust even for the difficult 10-20 ng/mL PSA range where standard PSA testing lacks discriminatory power.
The decisive features for identifying cancer among PI-RADS 3 lesions are far more subtle than those distinguishing PI-RADS 1 from PI-RADS 5. Standard deep learning models often fail because they were not designed to detect these fine-grained differences. The dual attention architecture of AttenNet specifically addresses this challenge.
Previous deep learning studies for prostate cancer have focused on patients across all PI-RADS categories, essentially treating all lesion ambiguity levels the same. This study is the first to build and validate a model exclusively for PI-RADS 3 lesions, recognizing that this group requires a fundamentally different approach.
Compared to radiomics approaches, which require slice-by-slice manual segmentation of lesions - a time-consuming bottleneck - AttenNet performs analysis automatically from the raw MRI volumes. This practical advantage makes it far more feasible for routine clinical deployment.
The use of biparametric MRI (without gadolinium contrast) also matters clinically. Eliminating contrast agents reduces examination time, cost, and the rare but real risk of gadolinium-related complications, while the soft attention module ensures the model appropriately weighs each sequence based on lesion location - the same logic used by radiologists.
AttenNet demonstrates that deep learning can meaningfully resolve the uncertainty inherent in PI-RADS category 3 findings. By accurately identifying which patients within this equivocal group have clinically significant cancer, the model functions as a clinical triage tool - deciding who needs immediate biopsy and who can be safely followed.
The multicenter external validation design - testing at three centers completely separate from training - provides strong evidence of generalizability. The consistent performance across centers with different patient populations, MRI scanners, and institutional practices strengthens confidence in real-world applicability.
Limitations include unequal sample sizes across centers and the retrospective design. Future prospective multicenter studies with larger patient numbers and standardized MRI protocols will be needed to fully establish clinical utility before AttenNet can be recommended for routine practice. Patients identified as low-risk will still require follow-up monitoring with repeat MRI or PSA.
As the burden of unnecessary biopsies continues to affect patient quality of life and healthcare resources, tools like AttenNet offer a path to smarter, more targeted use of prostate biopsy - reserving this invasive procedure for patients who truly need it while maintaining the ability to catch clinically significant cancers.