Deep Learning Image Reconstruction in CT: What It Holds for the Future

Author: Om Biju Panta, Beth Israel Deaconess Medical Center

Computed tomography has come a long way from filtered back projection, but the next major leap may be deep learning image reconstruction, or DLIR. In CT, DLIR is already showing that it can reduce noise, preserve diagnostic detail, and shorten reconstruction time, while supporting meaningful dose reduction. 

A Brief History: From Filtered Back Projection to Deep Learning

CT reconstruction began with filtered back projection, which was fast and mathematically simple but very sensitive to noise. Iterative reconstruction improved low-dose image quality, but often at the cost of unnatural, “plastic” image texture and longer processing times.

Deep learning reconstruction aims to combine the strengths of both methods. By learning from large datasets, it can suppress noise, preserve more natural image texture, and reconstruct images quickly, making it a promising tool for modern CT workflows.

What DLIR Actually Does and How

DLIR is not a single technique but a family of approaches, distinguished by where in the image formation pipeline the neural network is applied.

Direct DLIR bypasses traditional reconstruction entirely, generating images straight from the raw projection data (the sinogram). This approach can compensate for sparse or incomplete scan data and avoids introducing artifacts from FBP or IR.

Projection-space DLIR applies neural networks to the sinogram before conventional reconstruction takes place — correcting for beam hardening, scatter, and other artifacts at the source, which is more effective than trying to remove them afterward.

Image-space DLIR works after an initial FBP or IR reconstruction, using deep learning to refine and denoise the image. This is the most commercially prevalent approach because it is faster and easier to integrate into existing scanner platforms.

Hybrid DLIR combines projection- and image-space methods to balance quality and computational efficiency.

What the Evidence Shows

The clinical case for DLIR is now substantial. Studies consistently demonstrate:

• Noise reduction of up to 47% over iterative reconstruction, with contrast-to-noise ratio improvements of 92–94% in abdominal CT

• Radiation dose reductions of 30–71% compared to standard IR protocols while maintaining diagnostic image quality

• Natural image texture that resembles high-dose FBP — avoiding the plastic appearance that limited IR adoption

• Improved clinical performance across neuroimaging (gray-white matter differentiation), chest CT (pulmonary nodule detection), abdominal imaging (hepatic metastases, pancreatic tumors), cardiovascular CT (stent visualization, artifact reduction), and pediatric imaging where dose minimization is critical

A meta-analysis of the two most widely used DLIR algorithms found an average 39% reduction in image noise and a mean radiation dose reduction of 57–65% compared to IR methods.

The Challenges That Remain

The main concern is hallucination: DLIR models can create false features, especially near metal or at very low dose, and may also suppress small true structures. This is more concerning because the image often looks clean and convincing.

Training data diversity is another issue. Models trained on limited datasets may not perform well in rare anatomies, unusual pathology, or underrepresented patient groups.

DLIR is also still something of a black box, so unexpected artifacts can be hard to explain or trust. Explainable AI tools may help, including uncertainty maps and side-by-side comparisons.

Finally, vendor-specific reconstructions can look different from one system to another, which makes serial comparisons across scanners and institutions more difficult, especially in oncology follow-up. 

What the Future Holds

Photon-counting CT is the most important near-term development. By capturing individual X-ray photons with both energy and count data, it produces inherently higher-quality images — raising the ceiling for DLIR training data quality and creating reconstruction challenges that DLIR is uniquely suited to handle. The combination could produce a step-change in what CT imaging can achieve.

End-to-end models that map directly from raw projection data to diagnostic images — learning data correction, reconstruction, and noise reduction in a single unified network — represent a compelling longer-term direction. Some researchers are already exploring an even more radical step: AI-based diagnosis directly from sinogram data, without image reconstruction at all. Early feasibility studies have demonstrated detection of intracranial hemorrhage and vascular abnormalities from raw CT data — a capability that could transform emergency triage.

Finally, integration with downstream AI diagnostic tools will increasingly blur the line between producing an image and interpreting one. DLIR is not just a better reconstruction algorithm — it is the foundation layer of an AI-augmented CT workflow.

Conclusion

DLIR simultaneously improves image quality, reduces radiation dose, preserves natural image texture, and reconstructs quickly. That combination was not achievable before. The challenges that remain — hallucination risk, interpretability, data diversity — are real but tractable. As photon-counting CT hardware matures and DLIR algorithms grow more sophisticated, the CT scanner of the near future will see more clearly and expose patients to less radiation than anything that has come before.

References:

1.     Koetzier LR, Mastrodicasa D, Szczykutowicz TP, van der Werf NR, Wang AS, Sandfort V, van der Molen AJ, Fleischmann D, Willemink MJ. Deep learning image reconstruction for CT: technical principles and clinical prospects. Radiology. 2023 Jan 31;306(3):e221257.

2.     Sahu A, Mathur S, Takaoka H, Ota J, Meinel FG, Böttcher B, Magnin B, Jajodia A, Jensen CT. Transforming CT imaging with deep learning: Noise reduction, artifact management, and clinical applications–A comprehensive review. European Journal of Radiology Artificial Intelligence. 2025 Sep 18:100042.

3.     Mileto A, Yu L, Revels JW, Kamel S, Shehata MA, Ibarra-Rovira JJ, Wong VK, Roman-Colon AM, Lee JM, Elsayes KM, Jensen CT. State-of-the-art deep learning CT reconstruction algorithms in abdominal imaging. Radiographics. 2024 Nov 29;44(12):e240095.