Prostate cancer is the second most common cancer in men worldwide, with over 248,530 new cases diagnosed in the United States in 2023 alone. Early detection of clinically significant prostate cancer (csPCa) is essential for effective treatment planning, as it determines whether and how aggressively a patient needs to be treated.
Bi-parametric MRI (bpMRI) is a streamlined prostate imaging approach that uses only two key sequences: T2-weighted (T2w) imaging, which shows anatomical detail, and diffusion-weighted imaging (DWI), which reflects cellular density and water movement in tissues. Unlike full multiparametric MRI (mpMRI), bpMRI does not require contrast injection, making it faster, less expensive, and associated with fewer false positives -- while still maintaining adequate diagnostic accuracy for cancer detection.
The standardized MRI scoring system, PI-RADS v2.1, guides radiologists in assessing suspicious lesions. However, MRI interpretation is time-consuming, requires significant expertise, and shows considerable inter-observer variability -- especially at non-specialized centers. Radiologists can also make errors due to cognitive factors like mental fatigue, highlighting the need for computer-aided diagnosis (CAD) systems.
The prostate gland has an important anatomical structure: it is divided into the central gland (CG) and the peripheral zone (PZ). These zones differ in normal tissue appearance and in how cancer typically presents on MRI. Incorporating this zonal knowledge into AI models can help them focus their analysis on clinically relevant regions rather than processing the entire gland uniformly.
The core innovation of this study is a two-stage pipeline. The first stage segments the prostate zones (peripheral zone and central gland) to define the region of interest. The second stage then searches for cancer lesions within this anatomically-informed region. This approach contrasts with many prior methods that skip zone segmentation and attempt to detect cancer directly in the full image.
The rationale for staging is straightforward: a model that first knows where the prostate is and which zones are present can focus its cancer detection attention on diagnostically relevant regions. This reduces false positives from tissues outside the prostate and helps the model learn zone-specific cancer patterns, since cancer in the peripheral zone and transition zone looks different on MRI.
Both stages use an attention-guided U-Net architecture. The U-Net structure has an encoder that compresses the image into features and a decoder that reconstructs spatial maps for segmentation, with skip connections carrying fine spatial detail from encoder to decoder. Attention gates (AGs) are added to these skip connections to selectively emphasize the most relevant image regions and suppress irrelevant background, improving focus on anatomically significant structures.
The mathematical attention mechanism computes a weight for each spatial location based on both local features and a global context signal from deeper network layers. Features are then scaled by these attention weights before passing to the decoder, so the model dynamically learns to look at the right places rather than treating all pixels equally.
Stage 1 (prostate zone segmentation) was trained on a merged dataset of 443 samples from four public datasets: PROSTATEx (204 cases with zonal masks), Prostate158 (158 bpMRI cases), MSD Prostate (32 multiparametric cases from Radboud University), and a multisite dataset collected from six institutions providing six scanner and protocol variations. This diversity ensures the model generalizes across different MRI equipment and protocols.
Stage 2 (lesion segmentation) was trained on the PI-CAI challenge dataset, a large multinational cohort of 1,500 training cases from 9,129 patients across four European tertiary care centers. Among the training cases, 1,075 exhibited benign tissue or indolent cancer, while 425 were malignant cases -- labeled by either expert radiologists or a validated AI model. Five-fold cross-validation was used for rigorous evaluation.
All images were converted to a standardized NIfTI format, resampled to 0.5x0.5x3mm spatial resolution, and normalized to a 0-1 intensity range. Data augmentation included 20-degree rotation, horizontal flipping, and random noise addition (applied only to input images, not masks) to improve generalization and reduce overfitting.
Baseline comparisons included three pre-trained 3D CNN architectures with ImageNet backbones: VGG19, ResNet201, and SEResNet152. These represent well-established deep learning architectures commonly used in medical imaging, enabling a fair comparison between the proposed task-specific attention U-Net and generic pre-trained alternatives.
The attention-guided U-Net outperformed all baseline models for lesion segmentation, achieving a Dice Similarity Coefficient (DSC) of 0.82, an AUROC of 0.85, and an average precision of 0.80. This compares favorably to SEResNet152 (DSC 0.67), VGG19 (DSC 0.69), and ResNet201 (DSC 0.73), with all differences being statistically significant (p less than 0.05).
For prostate zone segmentation in Stage 1, the model achieved a DSC of 0.72 on the test set. Breaking this down by zone, the central gland was segmented more accurately (DSC 0.75, HD95 4.5mm) than the peripheral zone (DSC 0.70, HD95 5.2mm). The peripheral zone is anatomically thinner and more variable in appearance, making it inherently harder to delineate precisely.
An ablation study directly quantified the contribution of each component. A one-stage U-Net (skipping zone segmentation) achieved only DSC 0.63. Adding attention to the one-stage model improved this marginally to 0.64. A two-stage approach without attention gave DSC 0.70. The full two-stage attention model reached DSC 0.82 -- demonstrating that both the two-stage design and the attention mechanism are independently necessary for the best performance.
Compared to published literature, the model's DSC of 0.82 is below the 0.90 achieved by fully 3D CNN approaches (Saha et al.), but substantially above a recent 3D SegResNet approach that achieved only DSC 0.53 on bpMRI data. The model trades some accuracy for significantly reduced computational demands since it operates on 2D slices rather than full 3D volumes.
The key insight demonstrated by this study is that anatomical awareness improves AI performance on cancer detection. By first segmenting where the peripheral zone and central gland are, the lesion detection model can learn zone-specific cancer patterns. Cancer in the peripheral zone typically appears as a focal dark area on T2w imaging and restricted diffusion on ADC maps; cancer in the transition zone may look more heterogeneous and overlap with benign nodules.
The attention mechanism provides an additional layer of biological relevance. By dynamically weighting which spatial regions matter most at each scale of the network, the model naturally learns to focus on gland boundaries, lesion edges, and diffusion-restricted areas -- features radiologists themselves use to identify suspicious lesions.
The use of a 2D slice-wise architecture rather than a fully 3D model is a deliberate design choice that reduces GPU memory requirements and processing time, making the model more practical for deployment in clinical environments that may not have high-end computing infrastructure. The performance tradeoff -- marginally lower DSC than 3D models -- may be acceptable given the practical constraints of real-world radiology departments.
Training on multiple diverse datasets (different scanners, protocols, and patient populations) is an important step toward building models that work across the heterogeneity of real clinical practice, where images come from many different MRI systems with varying acquisition parameters.
This study demonstrates that integrating zonal anatomical knowledge into a deep learning pipeline for bpMRI prostate cancer segmentation produces meaningful improvements over methods that treat the whole image uniformly. The combination of two-stage processing and attention mechanisms is the key driver of performance gains.
The model's DSC of 0.82 achieves clinically useful performance while maintaining lower computational requirements than 3D alternatives. For clinical environments where access to high-performance computing is limited, this efficiency advantage could be decisive for actual deployment.
Key limitations include potential variation in performance across different MRI scanners and protocols beyond those represented in the training datasets. The model may also require fine-tuning for deployment in centers using different MRI field strengths or acquisition settings. Future work should focus on expanding training data diversity, testing additional augmentation strategies, and exploring hybrid 2D/3D architectures that balance accuracy with computational efficiency.