A-IDE : Agent-Integrated Denoising Experts
Keywords: Low-Dose CT, Image Denoising, LLM, Agents
TL;DR: Selecting specialized models via LLM agent boosts LDCT denoising quality.
Abstract: Recent advances in deep learning-based denoising methods have improved the quality of Low-Dose CT (LDCT) images. However, due to anatomical variability and limited data availability, a single model often struggles to generalize effectively across multiple anatomical regions. To address this limitation, we propose the Agent-Integrated Denoising Experts (A-IDE) framework. A-IDE integrates three region-specialized RED-CNN models under the control of a decision-making large language model (LLM) agent. This agent analyzes anatomical priors extracted from BiomedCLIP and dynamically routes incoming LDCT scans to the most suitable specialized model. We highlight three major advantages. First, A-IDE shows robust performance in heterogeneous and data-scarce environments. Second, the framework reduces risk of overfitting by distributing tasks among multiple experts. Finally, the fully automated agent-driven routing eliminates the need for manual intervention. Experimental results in the Mayo-2016 dataset confirm that A-IDE achieves superior performance in RMSE, PSNR, and SSIM compared to a single unified denoiser.
Submission Number: 17
Loading