Two-Stage Residual-Guided Diffusion Framework for High-Fidelity Pathology Image Restoration

Le résumé fourni par la source

The digitization of whole slide images in computational pathology frequently introduces out-of-focus blur, a prevalent artifact that can compromise diagnostic accuracy. Although recent diffusion-based restoration methods have shown notable success, they often fail to reconstruct the fine-grained cellular and textural details critical for histopathological interpretation. To address this limitation, we propose a two-stage residual-guided diffusion framework tailored for high-fidelity pathology image restoration. In the first stage, a latent diffusion process with a highly truncated diffusion schedule rapidly generates a coarse yet structurally coherent restoration. The second stage performs a targeted synthesis of the high-fidelity details by conditioning the diffusion process on the latent space residual between the coarse output and the ground truth. The experimental results demonstrate that the proposed method is superior to state-of-the-art image restoration models, yielding consistent improvements in key perceptual metrics such as LPIPS and CLIP-IQA. Qualitative evaluations further confirm that the proposed approach restores diagnostically critical cellular structures with markedly superior clarity, offering an effective solution for generating high-fidelity, degradation-free WSIs to enhance diagnostic accuracy in computational pathology.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.