Liu 2026 — ULTRA (ultrarapid cleared stimulated Raman 3D histology)
A Cell paper claiming FFPE-grade 3D histology of unfixed, unstained, unsectioned glioma tissue in 30 minutes, by stimulated Raman scattering plus unsupervised learning — filed from the abstract alone, so every number that would make the claim assessable is unread.
Read status: abstract only, and the gap matters more than usual here. This is a workflow claim — a time, a tissue volume, and an accuracy against conventional histology — and not one of the three is quantified in the abstract. What follows separates three things carefully: what the paper claims (its own abstract, quoted or closely paraphrased), what the bibliographic record shows that the abstract does not, and what is simply unknown. Nothing here is an appraisal of the work, because an appraisal is not available from 200 words. This folder is append-only, so if the full text is obtained later, the reading of it belongs on a wiki page, not in this file.
What the paper claims
ULTRA — “ultrarapid cleared stimulated Raman with AI”. The abstract describes a platform that is label-free, stain-free, fixation-free and section-free, built on the chemical specificity of stimulated Raman scattering (SRS) microscopy, and combining “the synergistic development of a one-step tissue-clearing protocol and unsupervised learning algorithms”. The stated output is “high-resolution, formalin-fixed, paraffin-embedded (FFPE)-grade deep 3D virtual histology within 30 min, covering orders of magnitude more tissue than slide-based methods”. On human surgical glioma samples it is said to resolve “key histological features in 3D” and to delineate “depth-dependent tumor infiltration margins at single-cell resolution”.
Two readings of that wording are worth fixing before they cause confusion:
- “FFPE-grade” is a quality comparison, not a process. The platform is claimed to be fixation-free; FFPE is named as the standard the image quality is measured against, not as a step in the protocol.
- “Deep” means depth into tissue, not deep learning. The claim is volumetric imaging some distance into an intact specimen; the learning component is separately described as unsupervised.
For a reader who has not met SRS: it produces image contrast from molecular vibrations — lipid-rich and protein-rich structures give different signal — so tissue architecture appears without any dye being applied. That is background, not from this paper; the paper’s own framing is only “the chemical specificity of SRS”. The route from raw SRS channels to something a pathologist reads as histology is exactly the part that is unread here.
What the record shows that the abstract does not
The indexed keywords name a diffusion model. The author keywords on the PubMed record are: 3D histology, diffusion model, gliomas, intraoperative histology, label free, section free, stimulated Raman scattering, tissue clearing. The phrase diffusion model appears nowhere in the abstract, which says only “unsupervised learning algorithms”. A generative model sitting between the microscope and the pathologist is a materially different proposition from unsupervised clustering or segmentation, because it raises the question of whether any part of the displayed histology is synthesised rather than measured. That is the single most important thing to settle from the full text, and it is invisible to anyone reading the abstract alone.
Where the work comes from. The author list is led by Fudan University groups — the Institutes of Biomedical Sciences with Shanghai Xuhui Central Hospital, and the Department of Physics with the Endoscopic Research Institute at Zhongshan Hospital — with clinical neurosurgery and pathology from Huashan Hospital, neuropathology from the Beijing Neurosurgical Institute, and one chemistry author at Columbia. Corresponding authors are Minbiao Ji and Lixue Shi. This is an optics- and chemistry-led programme with neuropathology attached rather than a pathology-informatics group, and the balance of the paper is likely to reflect that.
Access and reach. Closed access (Unpaywall reports closed with no OA location; Europe PMC reports inPMC: N), no preprint located through Crossref or Europe PMC, 106 references per Semantic Scholar / 94 linked by OpenAlex, no citations yet. Nothing in the open-access infrastructure will make this readable; a subscription or an author copy is the only route.
Prior conference disclosure, unread. A same-stem SPIE Photonics West presentation exists — “Ultrarapid deep 3D histology with cleared stimulated Raman imaging and unsupervised learning”, Lixue Shi, Advanced Chemical Microscopy for Life Science and Translational Medicine 2026, doi 10.1117/12.3083022, dated 2026-03-06, which falls between this paper’s revision and its acceptance. Only its existence is confirmed, from Crossref; the SPIE page returned no abstract text, so it has not been read and it corroborates nothing. It is recorded because it is the most likely place a shorter methods description sits in public.
The fields that are usually buried — here, all of them
AGENTS.md §8 asks for dataset, instrument, validation strategy and external validation on any paper describing a model. An unstated field is itself a finding; in this case the finding is uniform, and it is what makes the note short.
| Field | What is knowable without the full text |
|---|---|
| Specimens | “human surgical glioma samples” — no case count, no patient count, no grade or molecular-subtype breakdown, no statement of whether any specimen was diffuse versus circumscribed |
| Reference standard | Not stated. “FFPE-grade” implies a comparison against conventional histology of the same tissue, but neither the comparator nor who read it is described |
| Reader involvement | Not stated. Whether a neuropathologist ever read the 3D volumes, and under what design, is unknown — for an intraoperative claim this is the load-bearing experiment |
| What “accurately” means | No metric, no number, no dispersion anywhere in the abstract |
| What the 30 minutes covers | Not stated: clearing, imaging, reconstruction and inference are four different clocks, and the tissue volume they apply to is given only as “orders of magnitude more” than a slide |
| Resolution and depth | “single-cell resolution” and “deep” — no micrometres, no imaging depth, no volume in mm³ |
| The model | “Unsupervised learning algorithms”, plus a diffusion model keyword. Architecture, training data, and whether outputs are measured or generated: all unread |
| Validation strategy | Not stated |
| External validation | Not stated. Two independent neuropathology centres appear on the author list (Huashan, Beijing Neurosurgical Institute), which makes a multi-centre design possible; it is not evidence that there was one |
| Data and code availability | Not readable. Cell requires such a statement, so one exists in the paper |
| Prospective intraoperative use | The abstract claims the platform “enables more informed surgical decision-making”, not that any decision was made with it. Enabling language is not a use claim, and the difference should not be blurred in citing it |
What to check when the full text arrives
Written as a checklist because this note cannot be revised later.
- Is the displayed histology measured or generated? Where the diffusion model sits in the pipeline, what it is conditioned on, and whether a pathologist reads its output or the raw SRS channels. If it generates the H&E-like image, the failure mode is hallucinated structure at exactly the infiltrative margin the paper is about.
- The 30 minutes, itemised, against a stated tissue volume — and whether it includes the clearing step and the model inference or only the scan.
- How many specimens, from how many patients, and how they were selected.
- The reference standard and the reader design — matched conventional sections? how many readers? blinded? any agreement statistic?
- What “delineates infiltration margins” was measured against. Infiltrative margin has no crisp ground truth in glioma; whatever stands in for it determines what the claim means.
- Whether any cohort is external to the developing centre, and whether the Beijing neuropathology involvement is a second site or a second reader.
- Tissue viability afterwards — whether the specimen survives the protocol for routine histology, IHC and molecular work. For an intraoperative method used on tissue that must also be diagnosed conventionally, this decides whether it can ever be deployed.
- Data and code availability, and whether the clearing protocol is fully specified.
Why it earns space here
Not for anything it lets the group do — nothing in this department produces SRS images, and unlike a public dataset there is no artefact to reuse. It earns a note for two reasons.
It is the second route to 3D histology to arrive in this wiki within one month, and it attacks the opposite constraint. Gao 2026 — TRICARE (deep-learning triage of 3D pathology) takes cleared, dye-stained FFPE blocks imaged by open-top light-sheet over hours, and spends its effort on triage — choosing which of a thousand levels a pathologist should see. ULTRA claims to remove the staining, the fixation and most of the time, and to put the whole thing inside an intraoperative window. Whether the images are comparable is unknown from here, but the pair frames the question that matters for the modality: 3D pathology’s obstacle has been throughput and reading burden, and these two papers pick different halves of it.
It sharpens what “intraoperative” would have to mean. Frozen sections and smears exist because a 20-minute answer is worth a large loss of morphological quality. A method claiming FFPE-grade quality on that clock, on unfixed tissue, is claiming to remove a trade-off that every diagnostic protocol is built around — which is precisely why the unread numbers matter more than usual.
Related: Gao 2026 — TRICARE (deep-learning triage of 3D pathology) — the other 3D-histology route filed here, with full text read; light-sheet on cleared FFPE, and MIL triage of depth levels. Related: OTLS 3D Pathology Datasets — the public 3D-pathology data that exists, all of it from the light-sheet programme; this paper contributes none that is visible from the abstract. Related: Whole Slide Imaging — the substrate assumption both 3D routes remove: no slide, no section, and here no stain and no fixative either. Related: Intratumoral Heterogeneity — depth-dependent infiltration margin is a sampling claim about the same problem, measured in a volume instead of a section.