Prostate cancer is one of the most common cancers in men, and early detection significantly improves survival. Biparametric MRI (bpMRI) -- combining T2-weighted imaging and diffusion-weighted imaging (DWI) -- has emerged as the imaging standard for prostate cancer detection. Deep learning models trained on bpMRI can detect prostate lesions with high accuracy, but only when the test images closely match the type of images used during training.
A key challenge is that MRI acquisition protocols vary between clinical sites. Specifically, the strength of diffusion weighting in DWI is controlled by a parameter called the b value. Different hospitals use different b values depending on their preferences and scanner configurations. Although these protocol differences may seem small to a radiologist, they create significant differences in image appearance that confuse AI models trained on a different b-value standard.
This problem is known as domain shift: when a model encounters images from a distribution different from what it was trained on (called out-of-distribution or OOD data), its accuracy drops dramatically. In extreme cases in this study, a well-performing model dropped to an AUC of 0.49 -- essentially random performance -- when tested on images acquired with significantly different b values than the training set.
Existing solutions to domain shift either require labeled data from the new site (which defeats the purpose) or require training a separate model for every possible b-value combination (which is computationally impractical at scale). This paper proposes a practical, efficient solution called unsupervised domain adaptation (UDA) that requires no labels from the target site and uses a single unified model for all domain mappings.
The study assembled a large retrospective dataset of 5,150 male patients from nine different imaging centers across multiple countries, yielding 14,191 samples from 34 distinct combinations of b values. This scale and diversity is unusual in medical imaging AI research and was deliberately chosen to reflect the real-world variety of imaging protocols encountered in clinical practice.
All cases included T2-weighted imaging and DWI acquired at 1.5-Tesla or 3-Tesla scanners. Lesion annotations were based on clinical radiology reports and reviewed by an expert radiologist with 5 years of prostate MRI specialization. Each lesion was assigned a PI-RADS score -- the standard clinical scoring system for prostate lesion suspicion, ranging from 1 (very unlikely cancer) to 5 (very likely cancer).
For consistency, ADC maps and high-b-value images were recomputed at a fixed b value of 2000 sec/mm2 from each pair of low and high b-value images using a standardized mathematical formula, rather than using vendor-provided ADC maps that vary in their computation approach. This ensured comparable baseline measurements across sites before applying the domain adaptation step.
The dataset was split into 3,458 training cases and 1,692 independent test cases. The test set included samples from b-value combinations not represented in the training set, deliberately testing the model's ability to handle scenarios it had never seen during training -- the critical real-world challenge this study addresses.
The proposed framework has two components: a synthesis network and a detection network. The detection network (a U-Net with residual blocks) takes T2-weighted images, ADC maps, and DWI B-2000 images as input and generates a heatmap indicating where prostate lesions are located. This network was trained on the reference domain -- images acquired with b values matching the PI-RADS standard (low b value of 50 sec/mm2, high b value of 800 sec/mm2).
The synthesis network is an image-to-image translation model that converts images from any b-value combination (the target domain) into the visual style of the reference domain. This means that even if a hospital uses completely different b values, their images can be translated to look as if they were acquired with standard settings before being fed to the detection model. Critically, this translation happens without any labeled data from the target hospital.
The key innovation is the dynamic filter, a domain indicator component plugged into the synthesis network. Standard image translation models need separate networks for each domain pair. The dynamic filter allows a single unified generator to handle all domain mappings by accepting the actual b values as input metadata and dynamically adjusting its processing parameters accordingly. This makes the framework computationally efficient and easily extensible to new sites without retraining.
The framework is designed for plug-and-play use: once trained, the synthesis network can be placed upstream of any existing prostate cancer detection model to improve its performance on non-standard images. The detection model itself does not need to be modified or retrained, making this a practical add-on for clinical deployment.
Across all 1,692 test cases, the baseline supervised learning model achieved an AUC of 0.73 for PI-RADS 3 or higher lesions. The proposed UDA method improved this to AUC 0.79 -- a statistically significant improvement (p less than 0.001). For PI-RADS 4 or higher lesions, AUC improved from 0.77 to 0.80. These overall improvements demonstrate broad benefit across diverse imaging conditions.
The most striking results appeared in the most challenging scenario: the 361 test cases acquired with b values furthest from the standard (low b value of 150 or 200 sec/mm2, high b value of 2000 sec/mm2 -- Group 9). For this group, the baseline model performed at an AUC of 0.49 for PI-RADS 3 or higher -- essentially no better than random chance. The UDA method raised this to AUC 0.76, recovering near-normal detection performance from complete failure.
At the lesion detection level, the UDA method also reduced false positives per patient from 1.64 to 0.78 at a true-positive rate of 0.65, and from 2.31 to 1.11 at a true-positive rate of 0.70. This means not only were more cancers found, but far fewer benign areas were incorrectly flagged as suspicious -- a critical factor for clinical usability.
Importantly, the UDA method did not hurt performance on standard images. For images already acquired with reference domain b values (Group 4), performance was maintained. The generic model approach -- which simply retrains the model with more diverse b-value data -- actually harmed performance on standard images due to domain confusion, demonstrating that adding more training data alone is not the right solution.
The quality of generated images was measured using three technical metrics: peak signal-to-noise ratio (PSNR), mean square error (MSE), and structural similarity index measure (SSIM). For images originally acquired with b values of (150, 1500), generated images showed higher PSNR (32.89 vs 27.65), lower MSE, and higher SSIM compared to the original non-standard images -- confirming that the generated images more closely resemble reference-domain images in both overall brightness and structural detail.
A key visualization technique called t-SNE (t-distributed stochastic neighbor embedding) was used to project the internal representations of the detection model into a 2D plot. Original non-standard images formed a distinct cluster separated from reference-domain images -- confirming the domain shift problem. After passing through the UDA synthesis network, generated images formed a cluster that aligned closely with reference-domain images, confirming that domain adaptation was achieved at the level of what the detection model actually processes.
Qualitative image examples showed that for non-standard b-value images, generated images had visibly clearer lesion contrast compared to the originals, while detection heatmaps correctly highlighted lesion locations that the baseline model missed entirely. For standard images, no visible change was introduced by the synthesis step, confirming the framework's selectivity.
The study demonstrates a practical path to deploying AI prostate cancer detection at hospitals whose MRI protocols differ from the data used to train the model. Rather than requiring each hospital to re-label data or re-train the AI, the unified UDA framework can be applied once to any new imaging site using only unlabeled images -- a major reduction in the cost and effort required for clinical deployment. The framework is compatible with any existing pretrained detection model.
A key practical finding was that retraining the baseline model with more diverse b-value data (the generic model approach) was not a reliable solution. It produced unstable performance -- sometimes better, sometimes worse -- because the model lost specificity by trying to handle all b values simultaneously. This validates the UDA approach of keeping the detection model fixed and adapting images to it, rather than adapting the model to images.
Limitations include the current reliance on b values alone as domain metadata. Other sources of domain shift -- field strength, scanner manufacturer, number of signal averages, and T2-weighted image protocol -- were not incorporated. Some b-value groups had limited training samples (as few as 53 cases for Group 3), which contributed to weaker performance in those subgroups. Future work will extend domain metadata to include additional acquisition parameters and explore manifold learning for continuous domain representation.