For prostate cancer patients, PSMA PET/CT -- a highly sensitive molecular imaging scan that detects cancer by targeting the prostate-specific membrane antigen protein on cancer cells -- has become a powerful tool for identifying recurrence and guiding treatment decisions.
The challenge is that PET/CT combines two imaging technologies in a single scan: a PET scan that detects the radioactive tracer, and a CT scan used for attenuation correction -- a mathematical process that adjusts for how the body absorbs and scatters the PET signal. Roughly half the total radiation exposure from PET/CT comes from this CT component.
When patients require multiple follow-up scans during treatment, cumulative radiation exposure becomes a concern. Eliminating the CT portion of each scan would approximately halve the radiation dose per study, potentially allowing more frequent imaging or longer follow-up without exceeding safe exposure limits.
This study tested whether a deep learning AI model could perform attenuation correction directly from the raw PET data -- without needing a CT scan -- while preserving the quantitative measurements that oncologists rely on to track treatment response.
The AI model is based on a Pix-2-Pix generative adversarial network (GAN) -- an architecture specifically designed for image-to-image translation tasks. In this application, the network learns to translate non-attenuation-corrected PET images (NAC-PET) into attenuation-corrected equivalents (AC-PET), bypassing the need for an actual CT scan.
The architecture consists of two competing neural networks: a generator built on a U-net convolutional design that creates the synthetic corrected image, and a discriminator that tries to distinguish real corrected images from AI-generated ones. Training the two networks against each other forces the generator to produce increasingly realistic outputs.
Two normalization strategies were compared. Model V1 used standard SUV (standardized uptake value) scaling before training. Model V2 added a Nyul histogram normalization step designed to reduce variability in the raw PET signal, potentially improving consistency across different patients and scanning conditions.
The model was trained on 302 whole-body PSMA PET/CT scans from 283 prostate cancer patients at the National Institutes of Health, split into training (183 scans), validation (60 scans), and independent testing (59 scans) sets with strict patient-level separation to prevent data leakage.
In the independent test cohort, the AI-generated PET images achieved a median normalized mean squared error (NMSE) of 13.26%, mean absolute error of 3.59%, structural similarity index (SSIM) of 0.891, and peak signal-to-noise ratio (PSNR) of 26.82 decibels for the V1 model.
The V2 model with Nyul normalization produced slightly different results: a marginally higher NMSE (13.33%) but improved SSIM (0.951) and MAE (3.36%). The two models performed comparably overall, with each showing advantages on different metrics.
The largest pixel-level differences between AI-generated and real corrected images appeared in regions of very high tracer uptake -- specifically the kidneys, ureters, bladder, and salivary glands, where the PSMA tracer clears through normal biological pathways. These areas had uptake values that sometimes exceeded the model's 100 SUV preprocessing threshold, limiting correction accuracy in these normal anatomical structures.
Compared to previously published studies using FDG-PET (a different radiotracer) and synthetic CT generation, the SSIM and PSNR results from this PSMA-specific model were comparable, suggesting the approach generalizes reasonably across different imaging applications despite PSMA's challenging high-dynamic-range signal characteristics.
The clinically critical question was whether AI-generated images preserved the standardized uptake value (SUV) measurements used to quantify cancer activity in individual lesions. In the test cohort, 43 patients had 259 confirmed positive lesions distributed across lymph nodes (58%), bones (29%), and prostate (12%).
Statistical comparison of SUVmax and SUVmean -- the maximum and average tracer uptake within each lesion -- showed no significant differences between original AC-PET and AI-generated PET values (p = 0.33 and p = 0.18 for V1, p = 0.76 and p = 0.96 for V2). The AI preserved the quantitative cancer biomarkers that physicians use for clinical decisions.
Intraclass correlation coefficients (ICC) of 0.88-0.89 for both SUVmax and SUVmean confirmed high agreement between original and generated measurements -- comparable to the test-retest repeatability seen when the same patient undergoes two real PSMA PET scans in published repeatability studies.
Analysis of factors predicting AI error revealed a consistent pattern: lesions with higher uptake values had lower relative error, while denser lesions (higher Hounsfield unit values) and bone lesions had higher error compared to lymph node lesions. This suggests the model performs best for the high-uptake, soft-tissue lesions that are most common in metastatic prostate cancer.
The ability to generate attenuation-corrected PSMA PET images without a CT scan creates a practical pathway for reducing per-scan radiation exposure in patients requiring repeated follow-up imaging. This could allow more frequent monitoring during treatment without exceeding safe cumulative radiation limits.
A practical concern is that removing the CT component eliminates the anatomical roadmap that helps radiologists localize lesions precisely. The researchers propose that this can be addressed by registering the AI-corrected PET to a standard-of-care staging CT that most patients with advanced prostate cancer already undergo at treatment initiation -- providing anatomical context without requiring a dedicated attenuation correction CT for each follow-up scan.
The technology would be particularly valuable in clinical trial settings where patients may undergo a dozen or more scans over the course of treatment evaluation. Research protocols using PSMA PET for treatment response assessment currently face practical limits on scan frequency due to radiation concerns that this approach could substantially relax.
The model has not been evaluated in a prospective reader study where nuclear medicine physicians actually interpret the AI-generated images for clinical decisions. Validation of clinical diagnostic accuracy -- including assessment of whether any AI-generated artifacts could cause false positive or false negative lesion detections -- remains an important next step before clinical deployment.
The study used a single PSMA tracer (18F-DCFPyL) on a single scanner at one institution. While the controlled data collection ensures methodological consistency, it limits the model's generalizability to other PSMA tracers (such as 68Ga-PSMA-11, which is more widely used internationally), different scanner brands, and more diverse patient populations.
The model consistently underperformed in bone lesions compared to lymph node lesions, likely because dense bone tissue creates complex attenuation patterns that require accurate CT Hounsfield unit data to correct accurately. This is a meaningful limitation since bone metastases are common in advanced prostate cancer and are important targets for treatment monitoring.
An important future goal identified by the researchers is to extend the GAN to also generate a synthetic CT for anatomical localization purposes -- not just the attenuation-corrected PET signal. This would create a fully CT-free workflow without compromising the anatomical context needed for clinical interpretation.
The study did not assess the model's performance in patients with metal implants, ports, or prosthetics, which create CT artifacts that would influence attenuation correction in real imaging. Evaluating these challenging edge cases is important before widespread adoption, as these situations are common in the oncology patient population.