“The AoMP series produces original, academically rigorous content in physics, mathematics, signal processing, Fourier analysis, and MRI physics presented with clarity to make complex ideas accessible and illuminating.”

-- The Art of MRI Physics (AoMP)

Coming soon

Thanks for visiting the AoMP series!. Stay tuned for updates!

Work-in-progress

This is work-in-progress.

“Mystery creates wonder and wonder is the basis of man’s desire to understand.”

-- Neil Armstrong (1930–2012)

This is the first challenge in the AoMP series, in which I present a set of artifacts that I stumbled upon during my research.

TL;DR: In this challenge, you will learn about the classic artifacts in radial MRI. You will also encounter a so-called “mysterious” artifact. By solving this challenge, you will gain deep intuition about the physics behind these artifacts and how to mitigate them. The knowledge and intuition you gain are not only useful for radial MRI, but also for understanding MRI physics, other MRI artifacts, and signal processing and engineering more broadly.

If you have any suggestions or feedback, I would love to hear from you.

Radial MRI

Did you know that the first MRI image, acquired in 1973 by Paul Lauterbur, was actually radial? This pioneering work contributed to his receiving the Nobel Prize in Physiology or Medicine in 2003.

Although used to create the first ever MRI image, the radial MRI acquisition scheme is not commonly used in clinical practice today. Instead, the Cartesian acquisition scheme is more widely adopted in clinical MRI due to its simplicity, efficiency, and robustness to system imperfections. However, radial MRI has unique advantages such as reduced motion artifacts and applicability in certain applications, which has led to its continued use in research and specialized clinical settings.

With the improvement in hardware and reconstruction algorithms, radial MRI has been gaining more attention in recent years for its potential to provide high-quality images with reduced artifacts, especially in applications such as free-breathing, free-running 3D/4D/5D cardiac imaging and liver imaging. Radial MRI is perfectly suited for these applications because of its inherent robustness to motion artifacts as well as its ability to provide navigation information for motion correction and reconstruction.

Clinically, radial MRI has been used in liver MRI to reduce motion artifacts and capture the dynamic contrast enhancement of liver lesions. It also has been used in musculoskeletal MRI to generate CT-like contrast via ultrashort echo time (UTE) and zero echo time (ZTE) imaging.

Although not strictly radial, the PROPELLER/BLADE/JET MRI acquisition scheme, which samples k-space in a rotating fashion, has been widely used in brain and body MRI to reduce motion artifacts, where the artifacts are similar to those in radial MRI.

In this post, we will challenge you to identify the causes of the classic artifacts in radial MRI and propose solutions to mitigate them. During the process, we will also re-discover a mysterious artifact and strategy to mitigate it.

The “Halo” Artifact

The “Freaking” Artifact

How to Mitigate the “Freaking” Artifact?

The “Mysterious” Artifact

How to Mitigate the “Mysterious” Artifact?

Below is my summary of the study.

Study population

This retrospective study included 1,448 patients randomly selected from a pool of 29,114 patients underwent abdominal and pelvic CT scans between Jan 01, 2024 and Jun 30, 2024.

From 1,448 included patients:

• 300 patients were selected for optimizing the prompting of GPT-4o
• 600 patients were used for fine-tuning GPT-4o, in which 300 patients were used for training and 300 patients for validation
• 548 remaining patients were used as the test set to compare the performance of “prompting-only” GPT-4o, “fine-tuned” GPT-4o, and original human protocolers, including residents, fellows, and radiologists, who were originally selected the CT protocols.

Reference standard

Two subspecialty radiologists independently reviewed each case in the study cohort without knowledge of the original protocol selected. If the two radiologists agreed on the protocol selected, that protocol was used as the reference standard.

If there was disagreement between the two radiologists, the case was discussed with the third subspecialist expert radiologist to define the reference standard.

Model

GPT-4o (version 2024-08-06) was used as the base model. Model temperature was set at zero, which has better repeatability. All other parameters were kept at default values.

Prompt optimization

Relevant clinical information were given to GPT-4o in the form of detailed prompts. The prompt was iteratively optimized using the 300 cases reserved for prompt optimization.

Fine-tuning

GPT-4o was fine-tuned on 300 cases and then validated on another 300 cases.

Testing and evaluation

Testing was performed on 548 held-out internal cases. The model-selected and the original human-selected protocols were compared with the reference standard.

For protocols that did not match the reference standard, they were further classified as “equal alternative”, “reasonable but suboptimal”, or “inappropriate”.

Exact matches and “equal alternative” matches were considered optimal selections.

Results

Conclusion

GPT-4o is better than radiologists in selecting optimal abdominal and pelvic CT protocols.

Fine-tuning with labeled examples did not further improve performance beyond prompt optimization with detailed prompting instructions.

There were no significant differences in performance between residents, fellows, and attending radiologists.

References

Buckley BW, Dias AB, Deng Y, Schmidt H, Kielar A, Krishna S, Bhayana R. Optimizing Large Language Models for Automated Protocoling of Abdominal and Pelvic CT Scans: The Power of Context. Radiology. 2026 Jan 6;318(1):e252105. https://doi.org/10.1148/radiol.252105