CT-PrepAgent: Adaptive CT Data Preparation via LLM Agents

A recent preprint on arXiv (2608.01233) presents CT-PrepAgent, an innovative system designed for adaptive preparation of computed tomography (CT) data utilizing agents based on large language models (LLMs). This system tackles the issue of varied CT acquisitions and the differing requirements of downstream tasks, which hinder the adaptability of static data preparation processes. Current approaches depend on manually crafted or dataset-specific guidelines, resulting in a lack of flexibility in response to changes in acquisition conditions and analytical goals. CT-PrepAgent features a bounded policy and controlled deterministic execution for adaptive preparation. It creates structured data-task profiles through deterministic inspection, allowing a policy to select an appropriate DICOM series or preprocessing profile. The controlled execution flow manages, resolves, executes, and verifies the preparation stages. The study can be accessed on arXiv with the identifier 2608.01233.