EndoIR: Degradation-Agnostic All-in-One Endoscopic Image Restoration via Noise-Aware Routing Diffusion

TL;DR

EndoIR addresses the restoration of endoscopic images under multiple, unknown degradations by learning degradation-aware prompts from both spatial and frequency domains and by disentangling content from corruption cues through a dual-stream diffusion framework. Its key innovations—Dual-Domain Prompter, Task Adaptive Embedding, Dual-Stream Encoder, Rectified Fusion Block, and Noise-Aware Routing Block—enable efficient, degradation-agnostic denoising with improved preservation of anatomical details. The approach achieves state-of-the-art results on SegSTRONG-C and CEC, with fewer parameters and favorable downstream segmentation performance, highlighting potential clinical utility. This work demonstrates a practical path toward robust, real-time endoscopic image restoration that supports accurate diagnosis and surgical guidance in challenging visual conditions.

Abstract

Endoscopic images often suffer from diverse and co-occurring degradations such as low lighting, smoke, and bleeding, which obscure critical clinical details. Existing restoration methods are typically task-specific and often require prior knowledge of the degradation type, limiting their robustness in real-world clinical use. We propose EndoIR, an all-in-one, degradation-agnostic diffusion-based framework that restores multiple degradation types using a single model. EndoIR introduces a Dual-Domain Prompter that extracts joint spatial-frequency features, coupled with an adaptive embedding that encodes both shared and task-specific cues as conditioning for denoising. To mitigate feature confusion in conventional concatenation-based conditioning, we design a Dual-Stream Diffusion architecture that processes clean and degraded inputs separately, with a Rectified Fusion Block integrating them in a structured, degradation-aware manner. Furthermore, Noise-Aware Routing Block improves efficiency by dynamically selecting only noise-relevant features during denoising. Experiments on SegSTRONG-C and CEC datasets demonstrate that EndoIR achieves state-of-the-art performance across multiple degradation scenarios while using fewer parameters than strong baselines, and downstream segmentation experiments confirm its clinical utility.

Figures (5)

• Figure 1: Comparison of average restoration accuracy (PSNR/SSIM), model size, and inference speed (FPS, annotated above each marker). Both our EndoIR ($\gamma=0.5$) and ($\gamma=1$) achieve state‑of‑the‑art restoration quality, while the $\gamma=0.5$ variant offers a compact model size and high inference speed, making it well‑suited for real‑time clinical use.
• Figure 2: Overview of our EndoIR framework: Our denoising block consists of a Dual-Domain Prompter(DDP) to generate fine-grained prompt guidance; Task Adaptive Embedding(TAE) to dynamically learn task-specific conditions; Dual-Stream Encoder(DSE) to enable the dual-stream interaction for decoupled feature learning; Rectified Fusion Block(RFB) is designed to fuse dual-stream features; Noise-Aware Routing Block(NARB) for efficient feature refinement.
• Figure 3: The quantitative visualizations and error maps on the SegSTRONG-C ding2024segstrong dataset. Blue and red represent low and high error, respectively. (Zoom in to see details.)
• Figure 4: (a) and (b) shows t-SNE visualization between different tasks: (a) Output feature after passing Dual-Domain Prompter; (b) Task Adaptive Embedding output condition.
• Figure 5: Ablation study on Noise-Aware Routing Block under different ratio $\gamma$ on the SegSTRONG-C dataset.