Pathology image restoration recovering all-in-focus histology from single defocused focal planes, guided by semantic, defocus, and edge prompts.
No providers recorded yet. Browse all providers
Whole-slide scanners produce all-in-focus images the slow way: they capture each field of view at many focal depths and fuse the stack. Tissue is not flat, so a single focal plane leaves parts of the slide blurred, and the multi-plane workaround costs scanning time on every slide a laboratory digitizes. Computational restoration offers the alternative — scan one plane, recover the rest — but generic deblurring networks struggle on histology, where the blur pattern varies across the field of view and the densely packed nuclear texture pathologists read is exactly the high-frequency content those networks smooth away.
MoP, from ShanghaiTech University with collaborators at Shanghai Jiao Tong University and Nanjing University of Information Science and Technology, addresses this with a two-stage cascade and a deliberately heterogeneous set of conditioning signals. A transformer stage, P-former, restores fidelity; a diffusion stage, P-diffusion, restores perceptual sharpness. Both are steered by a "mixture of prompts": a defocus prompt describing the microscope optics, a semantic prompt read out of a frozen pathology foundation model, and an edge prompt from Canny extraction that constrains fine tissue structure.
The prompt mixture is what distinguishes MoP from generic restoration. Defocus alone says how blurred a region is but nothing about what tissue it contains; the pathology embedding supplies that context, and the edge map keeps the diffusion stage from inventing structure. Because a blurred input yields a degraded semantic embedding, MoP first restores the prompt itself, mapping the low-quality embedding toward the one a sharp image would produce. The work is a preprint and has not yet been peer reviewed.
P-former builds on a Restormer backbone, replacing its feed-forward blocks with a mixture-of-experts layer of three sub-experts gated by the input features concatenated with the pathology prompt. P-diffusion is built on ResShift and runs in four diffusion steps, with edge features fused into the conditioning path. Semantic prompts come from a frozen Prov-GigaPath encoder by default; substituting UNI or CONCH changes results marginally. Training used 100 histopathology slides from 100 patients covering multiple tissue types and organs, each scanned across 13 focal planes at 0.8 µm spacing, yielding 13,000 paired 256×256 patches, with 50 additional slides reserved to pretrain the defocus estimator. On the in-house test split, P-former reaches 29.14 dB PSNR and 0.8699 SSIM against Restormer's 28.85 and 0.8613, and the full cascade reaches 0.0735 LPIPS, the best perceptual score among all compared methods. On the public 3DHistech cervical cytopathology benchmark, where the defocus estimator and prompt restorer were transferred directly from the in-house data without fine-tuning, P-former reaches 33.68 dB PSNR and 0.903 SSIM and the cascade reaches 0.075 LPIPS, ahead of MPT+EFCR at 33.58, 0.887, and 0.119. Feeding restored patches to an abnormal-cell classifier lifts AUC from 90.90 on the raw defocused images to 96.58, approaching the 97.46 obtained from natively sharp scans.
The immediate use is slide digitization: a scanner that captures one focal plane and restores the rest cuts acquisition time per slide, which matters most for laboratories digitizing at volume and for cytology specimens where cells sit at varying depths. The downstream classification result matters more than the image metrics, since it shows restoration recovering diagnostic content rather than merely pleasant-looking texture — relevant to anyone running automated screening over archival or single-plane scans whose quality they do not control.
MoP demonstrates that pathology foundation model embeddings are useful as conditioning signals for low-level vision, not only as features for classification heads. Its ablations show the restored pathology prompt outperforming the raw low-quality embedding, the specific result that justifies the extra machinery. Only the scaffolding is public, though: the repository holds training code for the defocus encoder and the prompt restorer plus the curated abnormal-cell annotations for the 3DHistech test subset, while the README sections for training P-former and P-diffusion are empty and neither restoration model has code in the tree. No weights, no license, and no access to the in-house training slides accompany it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.