LLM Agent Automates CT Reconstruction Research
A new study is looking into how well a large language model (LLM) can carry out CT reconstruction research on its own. The LLM tweaks a solver step by step, runs cluster jobs, checks a specific metric, and then makes necessary changes. This metric is a calibrated headroom score compared to the FBP baseline within the visible area, using a shared differentiable fan-beam projector. The researchers tested 26 methods on low-dose CT from Mayo and a 128-view sparse breast task from the noiseless DL-Sparse-View Challenge. They evaluated selected iterations on a different test set and re-assessed each breast model with noisy inputs without retraining. The study aims to reduce manual work in comparing CT techniques and see if rankings from ideal data can predict real-world performance.
Key facts
- LLM agent autonomously edits solver, runs cluster jobs, reads metric, and revises
- Metric is calibrated headroom score against FBP baseline inside field of view
- All methods share same differentiable fan-beam projector
- 26 methods benchmarked on Mayo low-dose CT and DL-Sparse-View Challenge breast task
- Validation-selected iterations scored on held-out test set
- Breast models re-scored on noisy inputs (I_0 = 10^5 photons) without retraining
- Separate retraining on matched noise also performed
- Study compares rankings on ideal data vs. realistic noise
Entities
Institutions
- arXiv