A-IDE : Agent-Integrated Denoising Experts

TL;DR

Low-Dose CT denoising is hampered by anatomy-specific HU distributions and limited paired data. The authors propose A-IDE, an agentic framework that uses BiomedCLIP to semantically characterize scans and a GPT-4o-based agent to route each image to one of three region-specialized RED-CNN denoisers. The approach employs clustering to create anatomy-aware experts and an automated, graph-based decision pipeline to select the appropriate denoiser. On Mayo-2016, A-IDE achieves superior PSNR and SSIM while maintaining robustness across anatomies, illustrating the practical value of intelligent orchestration for LDCT denoising.

Abstract

Recent advances in deep-learning based denoising methods have improved Low-Dose CT image quality. However, due to distinct HU distributions and diverse anatomical characteristics, a single model often struggles to generalize across multiple anatomies. To address this limitation, we introduce \textbf{Agent-Integrated Denoising Experts (A-IDE)} framework, which integrates three anatomical region-specialized RED-CNN models under the management of decision-making LLM agent. The agent analyzes semantic cues from BiomedCLIP to dynamically route incoming LDCT scans to the most appropriate expert model. We highlight three major advantages of our approach. A-IDE excels in heterogeneous, data-scarce environments. The framework automatically prevents overfitting by distributing tasks among multiple experts. Finally, our LLM-driven agentic pipeline eliminates the need for manual interventions. Experimental evaluations on the Mayo-2016 dataset confirm that A-IDE achieves superior performance in RMSE, PSNR, and SSIM compared to a single unified denoiser.

Figures (10)

• Figure 1: Overview of the A-IDE (Agentic-Integrated Denoising Experts) Pipeline. The input CT image is first processed by BiomedCLIP to generate a semantic probability distribution over anatomical structures (e.g., lungs, spleen). This distribution is then passed to a LLM (gpt-4o) Agent along with textual descriptions of three specialized RED-CNN denoising models: Model 0, Model 1, Model 2. Based on the prompt specifying each model’s anatomical focus, the agent dynamically selects the most appropriate model for denoising. Finally, the chosen expert model reconstructs the denoised image patches and reports quantitative metrics: RMSE, PSNR, SSIM. Comparative evaluation shows that automatically routing images to a cluster-specific model improves denoising performance over a single baseline model.
• Figure 2: Comparison of 16 quarter 1 mm and full 1 mm patches preprocessed from Mayo-2016 Dataset.
• Figure 3: Distribution of anatomical structures in the Mayo-2016 Dataset based on image counts identified by highest similarity scores from BiomedCLIP embeddings. The x-axis indicates the number of representative structures. The y-axis represents the specific anatomical types.
• Figure 4: Five randomly selected images per cluster. Cluster 2 focuses on lungs, cluster 1 on pelvis, and cluster 0 on abdomial organs.
• Figure 5: Loss convergence for baseline model and three cluster-specific models. The x-axis indicates training iterations, and the y-axis represents training MSE loss.