Contrast-enhanced CT (CECT) is the standard imaging modality for detecting and characterizing renal cell carcinoma (RCC). The intravenous contrast agent causes enhancement of vascular tumors such as RCC, making them visible against the surrounding kidney parenchyma. Without contrast, many small renal masses are essentially invisible on CT.
However, CECT is not accessible to all patients. Individuals with compromised kidney function face the risk of contrast-induced nephropathy, a potentially serious complication. Others may have allergies to iodinated contrast agents. In settings with limited healthcare resources, the cost and availability of contrast agents can also be prohibitive. These barriers mean that a substantial subset of patients undergo plain CT without contrast, limiting the ability to screen for RCC.
This study explored whether a deep neural network (DNN) could synthesize realistic contrast-enhanced CT images from plain CT inputs, effectively converting non-contrast scans into diagnostically useful CECT equivalents. If successful, this approach would expand the population eligible for effective RCC screening without requiring actual contrast administration.
The DNN used in this study is based on the pix2pix framework, a conditional generative adversarial network (cGAN) designed for image-to-image translation tasks. In this context, pix2pix learns to map plain CT image patches to their corresponding CECT equivalents by training on paired image sets where both plain and contrast-enhanced scans were available for the same patients.
The training dataset comprised 99 patients who had both plain CT and CECT scans. The model learned the complex, spatially dependent relationship between tissue attenuation on plain CT and the expected contrast enhancement pattern, capturing differences between kidney cortex, medulla, tumor, and peritumoral fat. The validation set included 56 patients.
The generative adversarial approach involves two neural networks: a generator that produces synthetic CECT images and a discriminator that tries to distinguish synthetic from real images. Through adversarial training, the generator progressively improves until the discriminator can no longer reliably distinguish synthetic from real CECT, resulting in highly realistic synthetic images.
Image quality of synthetic CECT images was assessed using Zero-Normalized Cross-Correlation (ZNCC), a metric that measures the similarity between two images in a way that is insensitive to differences in mean intensity. A ZNCC above 0.70 indicates high structural and intensity concordance between the synthetic and real CECT images.
Importantly, ZNCC concordance above 0.70 was achieved for all tumor morphologies included in the study, including solid, partially cystic, and complex renal masses. This finding suggests that the DNN generalizes across the morphologic diversity of renal tumors rather than only performing well for a single tumor type.
Qualitative review of the synthetic images showed that contrast enhancement patterns in the renal cortex, medulla, and tumor tissue were visually plausible and anatomically consistent with what would be expected on true CECT. This provides confidence that the model is learning real physiologic enhancement relationships rather than superficially mimicking image texture.
Ten experienced urologists evaluated the synthetic CECT images and plain CT images in a blinded reading study, assessing each case for the likelihood of RCC. Using synthetic CECT, the urologists achieved an AUC of 0.892, compared to AUC 0.720 when reading plain CT alone. This difference was statistically significant (p less than 0.001), demonstrating that synthetic images meaningfully enhanced diagnostic accuracy.
Sensitivity was 72.7%, specificity was 75.7%, and overall accuracy was 74.2% using synthetic CECT. These performance metrics approach those reported for radiologist reading of true CECT in comparable patient populations, suggesting that synthetic enhancement provides clinically meaningful diagnostic value even if it does not perfectly replicate true contrast enhancement.
The fact that a panel of 10 urologists with varying levels of experience was used, rather than a single reader, strengthens the generalizability of the finding and reduces the risk that results reflect the exceptional skill of a single expert reader.
The authors note that this is the first study to use DNN-generated synthetic images specifically for RCC screening purposes. While image synthesis using GANs has been explored in other domains including brain MRI and cardiac imaging, application to renal CT for oncologic screening represents a novel contribution to the field.
One technical consideration is that the training data came from a single institution, and both plain CT and CECT scans were acquired using consistent protocols. Performance might vary if deployed across institutions with different CT scanner models, acquisition parameters, or contrast timing protocols. Multi-institutional training data would be needed to robustly address this limitation.
The pix2pix model is also inherently limited by the quality and diversity of training pairs. Rare tumor subtypes or unusual enhancement patterns not well represented in the training set may not be synthesized as accurately, a limitation that must be acknowledged when interpreting performance claims.
The primary clinical impact of this work lies in potentially enabling RCC screening for patients with contraindications to iodinated contrast. Patients with chronic kidney disease, contrast allergies, or limited access to contrast agents represent a population currently underserved by standard CT-based RCC detection.
If validated and deployed, synthetic CECT generation from plain CT could allow these patients to benefit from radiologic screening without the associated risks of actual contrast administration. This would be particularly valuable given that some patients with chronic kidney disease are at higher risk for RCC due to acquired cystic disease of the kidney.
Future work should validate performance on larger and more diverse datasets, compare against alternative non-contrast imaging strategies such as MRI, and assess whether the synthesis approach can be refined with advances in diffusion model-based image generation. Regulatory approval pathways for such AI-generated diagnostic images would also need to be carefully navigated.