High Fidelity Reconstruction for Low Dose Medical Imaging

This direction is split into other directions: low-dose imaging artifacts, low-dose CT image quality degradation. Its last version stays readable and citable here.

Version 1 of 1449 papers535 conceptsfrozen 2026-10-04

Medical imaging operates under a fundamental "Safety-Quality Paradox." To obtain the crystal-clear images required for a precise diagnosis, clinicians typically need high-energy radiation (in CT and PET) or powerful, expensive superconducting magnets (in MRI). Reducing the radiation dose to protect the patient or using portable, low-field magnets for bedside access results in images choked by quantum noise, streak artifacts, and poor spatial resolution.

The research direction emerging from these 251 papers focuses on breaking this trade-off. By moving beyond simple noise filters toward "physics-aware" deep learning and generative priors, researchers are now able to synthesize high-fidelity, diagnostic-quality images from data that was previously considered "clinically unusable."

From Denoising to Intelligent Synthesis

Early attempts to clean up low-dose scans relied on mathematical filters or iterative reconstruction. While these methods reduced noise, they often "smudged" fine anatomical details. The field shifted significantly when Hu Chen, Ge Wang, and colleagues demonstrated in Low-Dose CT via Deep Neural Network that convolutional neural networks (CNNs) could learn the complex mapping between noisy, low-dose patches and their clean, high-dose counterparts.

The intuition here is that noise in medical imaging isn't random white noise; it follows specific patterns dictated by the physics of the scanner. By training on "paired" data—scans of the same patient at both low and high doses—networks like the Stacked Competitive Network (SCN) developed by Wenchao Du can suppress artifacts while preserving the sharp edges of bones and lesions.

Incorporating the Physics of the Signal

The most recent breakthrough involves moving away from "black box" AI that only looks at pixels. Instead, researchers are embedding the physical laws of signal formation directly into the model architecture.

In the realm of portable MRI, Majida Ali Abed and Malek Al Abed introduced MRIQT, a framework that uses a physics-aware simulation to model how ultra-low-field signals are formed. Similarly, for PET imaging, Huidong Xie developed DDPET-3D, which uses "dose-aware embeddings." These models don't just guess what a clean image looks like; they use the known radiation dose or magnetic field strength as a guide, ensuring the final image is not just pretty, but quantitatively accurate.

Solving the Ground Truth Problem

A major hurdle in this field is that we cannot always scan a patient twice (once at low dose and once at high dose) to create training data. This "chicken-and-egg" problem has led to the rise of self-supervised and cyclic learning.

  • Cyclic Consistency: Peiyuan Jing and Zi Wang utilized the CSS-Diff framework, which uses a "cycle-consistent" constraint. The model learns to transform low-field to high-field and back again; if the round-trip doesn't match the original, the model corrects itself.
  • Hybrid Learning: Haoyang Pei and Ding Xia proposed a two-stage approach where the model first creates its own "pseudo-references" from noisy data, which are then used to train a more refined reconstruction network.

Generative Priors and Diffusion Models

The current state-of-the-art has moved toward Diffusion Models and Generative Adversarial Networks (GANs). Rather than just removing noise, these models treat reconstruction as a "denoising diffusion" process.

For example, the EASEL framework by Zhuonan He uses a score-based generative model to learn the "gradient" of what a healthy, high-dose image looks like. During reconstruction, the model iteratively nudges the noisy low-dose image toward this "ideal" manifold. This allows for extreme dose reductions—up to 200x in some PET studies by Junshen Xu and Enhao Gong—while maintaining the diagnostic utility needed to spot small tumors or subtle brain structures.

Clinical Implications

The ultimate goal of this research is democratization. By making 0.064T portable MRI scanners perform like 3T stationary giants, or making a 1% radiation PET scan as clear as a 100% dose scan, these technologies move high-end diagnostics out of specialized basements and directly to the patient's bedside, particularly for vulnerable populations like neonates and pediatric pneumonia patients.

Papers in this version

Showing 60 of 449 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_high_fidelity_reconstruction_for_low_dos_00adb231,
  title = {High Fidelity Reconstruction for Low Dose Medical Imaging},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-fidelity-reconstruction-for-low-dose-medical-imaging/rdr_00adb231fec04a269330ff3473686f5f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_high_fidelity_reconstruction_for_low_dos_00adb231_v1,
  title = {High Fidelity Reconstruction for Low Dose Medical Imaging},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-fidelity-reconstruction-for-low-dose-medical-imaging/rdr_00adb231fec04a269330ff3473686f5f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}