English | EN 
Introduction
In the hierarchy of medical imaging, the Chest X-ray (CXR) is the workhorse—fast, cheap, and accessible. However, it suffers from a fundamental physical limitation: it is a 2D projection of a 3D volume, leading to anatomical “superposition.” On the other end, Computed Tomography (CT) provides the gold standard for spatial resolution but at the cost of higher radiation and limited availability.
Recent advances in Generative AI have opened a provocative frontier: generating a Synthetic CT (sCT) directly from a 2D Chest X-ray. But a critical question remains: does this process actually add diagnostic value, or are we simply “hallucinating” anatomy?
The Value Proposition: What Does sCT Add?
If a synthetic CT cannot physically create information that wasn’t captured by the X-ray photons, why bother? The answer lies in Signal Unmixing.
- Spatial Disambiguation: By using Generative Adversarial Networks (GANs) or Diffusion Models trained on thousands of paired CXR/CT datasets, the AI learns to “unfold” superimposed structures. It can help determine if a suspicious opacity is located in the anterior or posterior segment of the lung—a task that is often ambiguous on a single frontal CXR.
- Feature Amplification: The AI can detect subtle texture variations (latent signals) invisible to the human eye and project them into a 3D-like cross-sectional view. This makes it easier for a clinician to characterize the margins of a nodule or the extent of a pleural effusion.
The Risks: Hallucinations and the “Mirror” Effect
The primary danger in Generative AI for medicine is the hallucination.
- Statistical Filling: If the AI is trained on a dataset where “Feature A” usually implies “Disease B,” it might generate “Disease B” on the sCT even if it’s not present in the patient, simply because it follows a statistical pattern.
- False Confidence: A synthetic CT looks like a real CT. This “visual authority” can mislead a radiologist into believing they are seeing a ground-truth 3D scan, when in fact they are looking at a probabilistic reconstruction.
Where is it Useless?
There are clear boundaries where sCT fails to provide any benefit:
- Acute Micro-Pathologies: A tiny, early-stage calcification that was physically blocked by a dense rib and left zero signal on the X-ray cannot be “recovered.” If the AI shows it, it’s a guess, not a diagnosis.
- Dynamic Changes: sCT cannot accurately reconstruct transient phenomena (like blood flow or specific contrast enhancement phases) that were not captured during the static X-ray exposure.
- High-Stakes Surgical Planning: You cannot use a synthetic CT for a lung resection. The risk of a 1-2mm spatial error inherent in generative synthesis is too high for surgical precision.
The Ideal Use Case: Triage and Low-Resource Settings
The real power of sCT is not to replace the CT scanner in a modern hospital, but to act as a Diagnostic Catalyst in underserved areas.
- Advanced Triage: In clinics where only X-ray is available, sCT can serve as a “super-filter,” helping doctors decide which patients urgently need a transfer to a center with a real CT.
- Education and Visualization: It serves as a powerful tool for explaining pathologies to patients or training junior residents to interpret 2D projections by looking at the synthesized 3D counterpart.
Conclusion: A “View” Rather Than a “Scan”
We must view Synthetic CT from CXR not as a new “modality,” but as an advanced visualization of a 2D signal. It is an interpretative aid that reveals latent information by leveraging the power of deep priors. As long as we treat the output as a probabilistic proposal rather than an absolute anatomical truth, we can safely use GenAI to bridge the gap between the 2D past and the 3D future of medical imaging.