The standard MRI protocol for diagnosing prostate cancer takes approximately 30 to 40 minutes and requires multiple imaging sequences. Among the most diagnostically valuable is diffusion-weighted imaging (DWI), which measures the random motion of water molecules in tissue. Cancer cells, which are densely packed, restrict water movement -- making them visible as areas of high signal. From DWI data, radiologists calculate a derived image called an Apparent Diffusion Coefficient (ADC) map, where cancer typically appears as dark (hypointense) regions.
ADC maps are used not only to detect prostate tumors but also to assess their aggressiveness, guide biopsy targeting, and monitor how a patient responds to treatments such as radiation or hormone therapy. Changes in ADC values over time provide a measurable window into whether the cancer is shrinking or growing in response to treatment -- making the ADC map one of the most clinically actionable images in the prostate MRI examination.
Despite their value, ADC maps are frequently compromised by image artifacts. The DWI sequences used to calculate ADC maps are particularly sensitive to patient motion -- even small involuntary movements such as breathing or bowel contractions -- as well as to distortions caused by rectal gas, which creates local magnetic field inhomogeneities. When artifacts degrade the DWI data, the resulting ADC map may be blurred, geometrically distorted, or misaligned, potentially causing radiologists to miss or mischaracterize lesions.
T2-weighted (T2W) MRI, by contrast, is far more robust and consistently produces high-quality images of the prostate anatomy. This study proposes a fundamentally different approach: instead of attempting to correct DWI artifacts after the fact, use generative AI to synthesize an ADC map directly from the T2W MRI image alone -- potentially eliminating the need to acquire DWI at all.
The AI-ADC model is built on a Generative Adversarial Network (GAN) framework, in which two neural networks compete: a generator that learns to produce synthetic ADC images from T2W MRI input, and a discriminator that learns to distinguish synthetic images from real ones. This adversarial training pushes the generator to produce increasingly realistic ADC maps. The generator is based on a ResNet architecture with nine residual blocks, chosen for its efficiency and fast inference speed.
A key technical innovation is the incorporation of a Convolutional Block Attention Module (CBAM), which combines two types of attention: channel attention, which identifies which feature maps (channels) contain the most relevant information, and spatial attention, which identifies which spatial locations in the image are most informative for the task. Together, these attention mechanisms direct the network to focus computational resources on the prostate gland and abnormal regions rather than distributing equal attention across irrelevant background tissue.
A second major innovation is the use of contrastive learning via patch-wise noise contrastive estimation loss. Rather than requiring perfectly paired T2W and ADC scans from the same patient at the same time -- a strict requirement that limits training data -- AI-ADC uses unpaired images and learns by comparing image patches: patches at the same spatial location in the T2W input and generated ADC output should be similar (positive pairs), while patches from different locations should differ (negative pairs). This allows the model to preserve structural details and spatial relationships without requiring matched pairs.
An additional self-regularization loss penalizes the model if the generated ADC map deviates too far from the original T2W structural content, ensuring that anatomical boundaries such as the prostate capsule and prostatic zones are accurately preserved in the synthetic image. The full training loss combines the adversarial GAN loss, the contrastive loss, and this self-regularization term, each weighted by adjustable parameters.
The model was developed using MRI data from the National Cancer Institute (NIH), where 506 patients undergoing prostate MRI for suspected or confirmed cancer were included in training. All scans were acquired on a 3 Tesla MRI system and were evaluated by an expert radiologist with 20 years of experience using criteria matching the Prostate Imaging Quality (PI-QUAL) grading system. Scans were included if they provided sufficient diagnostic information despite the presence of some artifacts; 399 had acceptable quality and 106 had high quality.
Images were preprocessed by cropping the prostate gland out of the full MRI scan using manually drawn segmentation masks, focusing the model on the relevant anatomy. Prostate-specific cropping was essential: experiments without cropping failed because the model was distracted by variable anatomy outside the prostate, demonstrating that boundary delineation is a prerequisite for successful ADC synthesis.
External validation was performed on the publicly available ProstateX dataset, collected at Radboud University Medical Center in the Netherlands as part of a global challenge held in 2016 to 2017. This dataset comprises 344 patients scanned on different Siemens scanner hardware, at a different institution, with somewhat different imaging protocols than the NIH training data. Testing on ProstateX assesses how well the model generalizes beyond the single institution where it was developed.
AI-ADC was benchmarked against three established image translation methods: CycleGAN, CUT (Contrastive Unpaired Translation), and StyTr2, a transformer-based style transfer model. On the NIH internal test set of 195 patients, AI-ADC achieved the highest Structural Similarity Index (SSIM) of 0.863 and the lowest Frechet Inception Distance (FID) of 31.992, compared to 0.855 and 43.458 for CycleGAN, 0.797 and 179.983 for CUT, and 0.824 and 58.784 for StyTr2. Higher SSIM indicates closer structural similarity to the real ADC maps; lower FID indicates the distribution of generated images more closely matches real ADC images.
On the external ProstateX validation set of 147 patients, AI-ADC similarly achieved the best SSIM of 0.647 and the lowest FID of 113.876, with the next-best model (StyTr2) scoring 0.534 SSIM and 120.393 FID, and CUT performing worst at 0.467 SSIM and 224.678 FID. The absolute SSIM values are lower on ProstateX because those scans were acquired on older hardware (circa 2010), but the relative performance advantage of AI-ADC held consistently.
Critically, AI-ADC demonstrated resistance to artifacts that typically degrade clinical ADC maps. In cases where the original ADC map was severely compromised by rectal gas distortion, the AI-generated map correctly delineated the posterior prostate boundary and identified intraprostatic zones that were invisible on the real ADC map. In one case involving a UroLift device -- a metal implant that creates severe susceptibility artifacts in DWI -- the AI-ADC map showed clear intra- and extraprostatic structures with minimal artifact, a result that would be clinically impossible using conventional DWI-based ADC calculation.
Histopathologic validation was performed on one case with a PI-RADS 5 lesion confirmed by radical prostatectomy pathology. The hypointense region on the AI-generated ADC map aligned precisely with the tumor location confirmed in the whole-mount histology specimen, providing direct evidence that the AI-ADC map reflects real tumor biology rather than arbitrary image pattern matching.
The most immediate clinical application is eliminating or shortening the DWI acquisition during prostate MRI. Because DWI is the sequence most prone to artifacts and requires additional scanner time, generating a synthetic ADC map from the already-acquired T2W MRI could reduce total scan time while simultaneously providing a cleaner, artifact-free image. This would make prostate MRI faster, less expensive, and more accessible as a routine screening tool.
For patients where DWI is technically compromised -- due to bowel motion, rectal gas, metal implants such as hip prostheses or UroLift devices, or patient restlessness -- the AI-generated ADC map could rescue the diagnostic exam entirely. Currently, radiologists reviewing such scans must either accept a degraded ADC map or repeat the examination, both of which are costly and clinically unsatisfying. AI-ADC offers a third option: synthesize a high-quality ADC image from the clean T2W scan that was acquired simultaneously.
The model's inference time of approximately 6.5 milliseconds per MRI slice makes it suitable for real-time clinical workflows. This stands in contrast to CycleGAN, which requires 17.7 milliseconds per slice -- still fast in absolute terms, but nearly three times slower than AI-ADC. For the typical prostate MRI with dozens of axial slices, this difference translates to meaningful workflow efficiency.
Future steps planned by the research team include a multi-reader study, in which multiple radiologists will evaluate AI-generated ADC maps alongside real ones in a blinded setting to assess clinical acceptability and diagnostic accuracy. The team also plans to formally test whether AI-ADC improves lesion detection rates compared to real ADC maps for PI-RADS-graded lesions, which is the most direct measure of clinical value.
The most significant limitation is that AI-ADC was developed at a single institution (NIH) using one scanner platform and one acquisition protocol. While the ProstateX results are encouraging, they involve older hardware and the SSIM values drop substantially compared to the internal test set, indicating that scanner and protocol differences affect synthesis quality. True clinical validation requires prospective, multi-center testing across diverse scanners, field strengths, and patient populations before the technology can be widely deployed.
The model also currently depends on a prostate segmentation mask to crop the gland before synthesis. While an automated in-house segmentation model was shown to produce adequate masks (mean Dice score of 0.78), removing this preprocessing dependency entirely would simplify clinical implementation. The failure mode analysis showed clearly that when the full MRI without cropping was fed to the model, synthesis quality collapsed due to the variability in prostate size and position across different patients.
The study also did not formally measure whether the AI-ADC maps improve lesion detection performance compared to real ADC maps -- the most clinically meaningful question. All quantitative results are based on image similarity metrics (SSIM, PSNR, FID) that measure how closely the generated images match real ones, but these do not directly tell us whether a radiologist reading the AI map would make the same or better clinical decisions than one reading the original ADC map.
Looking forward, the researchers discuss the potential of integrating PSA density -- a blood-based biomarker known to predict biopsy outcomes -- as an additional input alongside T2W MRI to improve ADC synthesis. They also note interest in diffusion model-based generative approaches, which have shown impressive image synthesis results in recent years, though their substantially higher computational cost currently limits real-time deployment. As hardware capabilities grow and diffusion models become more efficient, they may offer a future pathway to even higher fidelity synthetic imaging.