Deep Learning Synthesizes Contrast-Enhanced CT and MRI Scans

The Clinical Imperative for Contrast-Free Diagnostic Imaging

Contrast-enhanced imaging remains the cornerstone for the precise characterization of malignancies, ranging from hepatocellular carcinoma to head and neck squamous cell carcinoma [1, 2]. However, the administration of iodinated and gadolinium-based agents carries inherent risks of adverse reactions and systemic complications, particularly nephrotoxicity in patients with impaired renal function. While artificial intelligence has already demonstrated high sensitivity in detecting lesions across various modalities, most current models still rely on traditional contrast-enhanced sequences to achieve diagnostic precision [3, 4, 5]. As clinicians seek to minimize procedural risks while maintaining high diagnostic standards, the development of virtual enhancement techniques has become a priority. A recently published scoping review now evaluates the current state of deep learning models designed to synthesize these critical diagnostic images without the need for chemical contrast injection.

Mapping the Landscape of Synthetic Contrast Research

To evaluate the current state of virtual enhancement, the researchers conducted a comprehensive scoping review following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) guidelines, which is a standardized framework designed to ensure transparent and complete reporting of evidence syntheses. The authors performed an exhaustive search of PubMed, Embase, Scopus, and Web of Science from the inception of each database through September 2025. This systematic search identified 56 studies that met the inclusion criteria, providing a robust dataset for analyzing how deep learning models synthesize contrast-enhanced images. The distribution of research was relatively balanced across imaging modalities, consisting of 25 CT studies and 31 MRI studies. The clinical focus of these investigations was largely concentrated on specific anatomical regions where contrast is traditionally vital for lesion characterization. The brain was the most frequent target for synthetic contrast imaging, followed by applications in the head and neck, breast, and liver. These areas represent high-volume clinical priorities where reducing the cumulative load of iodinated or gadolinium-based agents could significantly impact long-term patient safety, particularly for patients requiring longitudinal surveillance. To generate these synthetic images, the studies employed two distinct data input strategies. The majority of the research, representing 71 percent of the included studies, used non-contrast inputs to generate synthetic images, meaning the models attempted to predict enhancement patterns using only the native tissue characteristics found on unenhanced scans. The remaining 29 percent of studies used modified-contrast strategies, which involve using a low-dose or suboptimal contrast injection as the baseline for the model to synthesize a full-dose equivalent image, potentially offering a middle ground for dose reduction.

Evolution of Deep Learning Architectures and Data Scale

The technical foundation of synthetic contrast imaging has evolved rapidly, moving from established frameworks to more sophisticated architectures. Throughout the reviewed literature, generative adversarial networks (a class of machine learning frameworks where two neural networks compete to create realistic data) were the predominant model class used to synthesize images. These networks function by having one model generate a synthetic image while a second model attempts to distinguish it from a real contrast-enhanced scan, a process that iteratively improves the visual fidelity of the output. More recently, the field has seen the emergence of diffusion models and transformer models (advanced architectures that handle complex data patterns by focusing on global relationships within the image), which began to appear in the literature after 2023. These newer architectures are particularly adept at capturing the intricate spatial relationships and subtle texture variations required to accurately simulate the distribution of contrast agents in heterogeneous tissues. Despite the increasing complexity of these models, the scale and diversity of the data used to train them remain significant hurdles for clinical implementation. The dataset sizes in the reviewed studies ranged from 10 to 7306 subjects, with a median of 218, suggesting that many models are developed on relatively small cohorts that may not capture the full spectrum of pathological or anatomical variability. Furthermore, 57 percent of the studies were conducted at a single center, a factor that limits the external validity of the findings. For the practicing clinician, this concentration of single-center data means that a model's performance may be optimized for specific scanner hardware or institutional imaging protocols, potentially leading to reduced accuracy when applied to different patient populations or imaging environments.

Quantitative Fidelity and Perceptual Quality Metrics

The technical validation of synthetic imaging relies heavily on mathematical benchmarks that compare the generated output to a ground-truth, contrast-enhanced reference scan. According to the review, quantitative fidelity (the mathematical similarity between synthetic and real images) was evaluated in 88 percent of studies, indicating a strong research focus on pixel-level accuracy. To quantify this accuracy, researchers primarily utilized the structural similarity index, which measures how well the model preserves the luminance, contrast, and structure of the original anatomy. The reported structural similarity index values ranged from 0.73 to 0.99, where a value of 1.0 represents a perfect mathematical match to the reference image. For the clinician, these high values suggest that the deep learning models are highly effective at replicating the expected anatomical appearance of contrast-enhanced tissues. Beyond structural alignment, the studies also measured the clarity of the synthetic signal using peak signal-to-noise ratios (a measure of signal strength relative to background noise). The peak signal-to-noise ratios ranged from 22 to 51 dB, with higher decibel levels indicating that the synthetic contrast enhancement is clearly distinguishable from the underlying image noise. This metric is particularly relevant for identifying small or low-contrast lesions that might otherwise be obscured by artifacts. Collectively, the data indicate that deep learning-based synthetic contrast imaging shows high quantitative and perceptual fidelity, producing images that are often visually indistinguishable from those acquired using traditional contrast agents. However, the researchers emphasized that while these mathematical metrics are robust, they do not inherently guarantee that the synthetic images are suitable for making definitive clinical diagnoses.

Bridging the Gap to Diagnostic Interchangeability

The clinical utility of synthetic imaging depends less on mathematical pixel matching and more on its ability to replicate the diagnostic accuracy of traditional contrast-enhanced scans. According to the review, diagnostic performance (the ability of the synthetic images to correctly identify or exclude disease) was assessed in only 30 percent of the studies. Within this subset of research, the reported sensitivities for synthetic contrast images ranged from 72 percent to 92 percent, while reported specificities ranged from 59 percent to 95 percent. These wide ranges suggest that while some models effectively identify pathology, others may carry a higher risk of false positives or missed findings compared to the current gold standard of physical contrast administration. Beyond binary diagnosis, the researchers examined how these images function in a clinical workflow. Radiologist-rated image quality (a subjective evaluation of whether the synthetic scan meets the necessary standards for clinical interpretation) was evaluated in 54 percent of the studies. While these ratings were generally favorable, they do not always correlate with objective diagnostic success. To address this, some researchers looked at task-based performance (the effectiveness of the images in specific clinical applications, such as tumor volume measurement or lesion segmentation), which was assessed in 39 percent of the studies. These task-specific evaluations are critical for clinicians who rely on contrast to define surgical margins or monitor treatment response in oncology. Despite the high quantitative fidelity reported across the literature, the transition from research to the radiology suite faces significant hurdles. The researchers concluded that evidence supporting diagnostic interchangeability (the capacity to use synthetic images in place of traditional contrast-enhanced scans without compromising clinical accuracy) and routine clinical use remains limited. For the practicing physician, these findings indicate that while deep learning can generate visually convincing contrast effects, the technology has not yet reached the level of validation required to safely replace iodinated or gadolinium-based agents in standard diagnostic protocols.

References

1. Elhaie M, Koozari A, Arjmandi M, Najafizade N. Deep learning for hepatocellular carcinoma segmentation in MRI: A systematic review of models, performance, and challenges.. Medicine. 2025. doi:10.1097/MD.0000000000047061

2. Johnson DE, Burtness B, Leemans CR, Lui VWY, Bauman JE, Grandis JR. Head and neck squamous cell carcinoma. Nature Reviews Disease Primers. 2020. doi:10.1038/s41572-020-00224-3

3. Rokhshad R, Salehi SN, Yavari A, et al. Deep learning for diagnosis of head and neck cancers through radiographic data: a systematic review and meta-analysis.. Oral radiology. 2024. doi:10.1007/s11282-023-00715-5

4. Xue P, Wang J, Qin D, et al. Deep learning in image-based breast and cervical cancer detection: a systematic review and meta-analysis. npj Digital Medicine. 2022. doi:10.1038/s41746-022-00559-z

5. Abdullah KA, Marziali S, Nanaa M, Sánchez LE, Payne NR, Gilbert FJ. Deep learning-based breast cancer diagnosis in breast MRI: systematic review and meta-analysis.. European radiology. 2025. doi:10.1007/s00330-025-11406-6