Diffusion super-resolution for H&E whole-slide images, steering denoising with SAM-derived structural anchors and adapted DINOv3 semantic control.
No providers recorded yet. Browse all providers
Digital pathology assumes high-resolution whole-slide images, but storage cost, bandwidth, and limited access to high-magnification scanners mean lower-resolution slides are what many clinical workflows actually produce. Super-resolution can recover diagnostic detail without new hardware, yet the two dominant families fail in characteristic ways: deterministic regression models minimize pixel error and smooth away the nuclear and stromal texture pathologists read, while GAN and diffusion models synthesize convincing high-frequency content that need not correspond to the tissue on the slide.
S³-Diff, from a team led by South China University of Technology with collaborators at Guangdong University of Technology and the Affiliated Cancer Hospital of Guangzhou Medical University, targets that mismatch. It constrains a latent diffusion denoiser with two training-time signals: Specimen-aware Structural Anchoring (SSA), which builds a specimen-specific structural reference from a frozen Segment Anything backbone and the gradient discrepancy between the high- and low-resolution images, and Structure-guided Semantic Fidelity Tuning (SSFT), which adapts DINOv3 representations under that supervision into a control signal computable from the low-resolution input alone.
The entry point differs from the tile-level representation models that dominate computational pathology — UNI and Virchow learn features from slides already scanned at full resolution, whereas S³-Diff addresses whether a usable slide exists at all. Its closest analogue is restoration work such as UniFMIR in fluorescence microscopy, though S³-Diff also scores reconstructions by a clinical endpoint rather than image metrics alone.
S³-Diff builds on a frozen Stable Diffusion v1.5 latent backbone. The U-Net and VAE stay fixed; training updates only a ControlNet, an energy head, DINOv3's final transformer block, and its LayerNorm parameters, using AdamW over 60,000 micro-batch iterations at an effective batch size of two, with learning rates of 3 × 10⁻⁶ for the control branch and 6 × 10⁻⁷ for the adapted DINOv3 parameters. Training used 150 TCGA whole-slide images — 50 each from lung (LUAD), kidney (KIRC), and liver (LIHC) cohorts — split 7:1:2 by patient, with 512 × 512 patches bicubically downsampled to 128 × 128 for 4× super-resolution, and the checkpoint chosen by lowest validation LPIPS.
Against deterministic and generative baselines including SwinIR, SHISRCNet, ESRGAN, SinSR, STAR-RL, UPSR, and SuperDiff, S³-Diff records the best LPIPS and ST-LPIPS on all three TCGA cohorts (LPIPS 0.1651, 0.1982, and 0.1852 on LUAD, KIRC, and LIHC), cutting LPIPS relative to SwinIR by 31.89%, 10.07%, and 15.89%. On the held-out SurGen colorectal cohort of 46 patients it achieves the best ST-LPIPS (0.1711) and Grad-L1 (0.0032), with LPIPS within 1.28% of the best result. An ABMIL survival model trained on its reconstructions reaches a concordance index of 0.8044, above bicubic interpolation at 0.7241 and the high-resolution reference at 0.7843 (log-rank p = 6.25 × 10⁻⁵).
The immediate use is recovering diagnostic detail from slides never captured at high magnification: retrospective archives digitized at low objective settings, laboratories without high-magnification scanners, and pipelines where storage or bandwidth constraints force downsampling. Because reconstructions preserve survival-relevant signal, they can feed weakly supervised multiple-instance learning pipelines for outcome prediction rather than serving only as viewing aids — useful for retrospective studies where slide quality is uneven across contributing sites.
S³-Diff is a preprint awaiting peer review, and no source code or weights have been released; the authors state that code will be made public. The evaluation covers three TCGA cancer types and one external cohort under a single 4× bicubic degradation model, so behavior under other magnification ratios and real acquisition conditions is untested, and the authors identify multicenter evaluation as future work. They are explicit that the model is for research, cannot replace pathologist review without prospective validation, and that its outputs must be identified as computational reconstructions. The more transferable contribution is methodological: tying restoration supervision to a prognostic endpoint, and validating on a cohort withheld from checkpoint selection, sets a more demanding bar than the perceptual metrics on which pathology super-resolution is usually judged.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.