One of the major challenges in glioma surgery is that tumor boundaries are often difficult to define. Diffuse gliomas can infiltrate surrounding brain tissue beyond what surgeons can see directly or what conventional imaging can fully resolve. During surgery, neurosurgeons must remove as much tumor as possible while preserving critical brain functions. This makes accurate intraoperative assessment of tumor infiltration essential for more precise surgical decision-making.
Current intraoperative pathology largely relies on frozen section analysis. Although this approach can provide rapid pathological information, it is typically based on a limited number of two-dimensional tissue sections and may be affected by sampling location, section thickness, freezing artifacts, and staining quality. By contrast, three-dimensional histology preserves spatial tissue architecture and allows tumor morphology and cellular distribution to be examined across a tissue volume. However, existing 3D histology methods often require complex tissue processing, labeling, staining, and prolonged imaging, taking hours or even days and making them difficult to use during surgery.
A study published in Cell now reports a new platform called ULTRA, short for ultrarapid cleared stimulated Raman with AI, designed to address this gap. Developed by researchers from Fudan University, Huashan Hospital of Fudan University, Zhongshan Hospital of Fudan University, Beijing Neurosurgical Institute, Capital Medical University, and collaborators, ULTRA combines rapid tissue clearing, stimulated Raman scattering microscopy, and AI-based virtual staining to perform deep 3D histological imaging of surgical tissue without conventional staining or sectioning.
The core advance of ULTRA is its ability to compress deep 3D histological analysis into an approximately 30-minute workflow. The researchers first developed a rapid tissue-clearing method compatible with stimulated Raman scattering microscopy, allowing fresh or fixed brain tissue to become transparent enough for millimeter-scale volumetric imaging. The tissue is then imaged by stimulated Raman scattering microscopy, which captures intrinsic chemical-bond vibrational signals without external labels. Finally, AI algorithms reconstruct and convert the imaging data into H&E-like 3D virtual histology, presenting tissue morphology in a format closer to what pathologists are familiar with.
“Conventional intraoperative pathology often relies on a limited number of two-dimensional sections, while tumors themselves are highly heterogeneous three-dimensional entities. The goal of ULTRA is to obtain a more complete 3D histological view within a timeframe close to intraoperative pathology, without destroying the tissue. We hope this technology will help doctors better understand glioma infiltration margins in the future and provide more tissue-level information for intraoperative decision-making,” said Dr. Lixue Shi of Fudan University.
Technically, the ULTRA staining pipeline consists of three sequential AI modules. The first is a diffusion-model-based depth-aware denoising module, which restores image quality as signal declines with imaging depth. The second uses a conditional generative adversarial network to predict the protein channel from the lipid channel, reducing the time and instrumentation burden associated with dual-channel volumetric acquisition. The third uses a virtual H&E staining model to convert stimulated Raman imaging data into H&E-like images that are more interpretable for pathologists. In this study, AI is therefore not used simply to “replace” diagnostic judgment, but to support image restoration, channel prediction, virtual staining, and 3D margin visualization.
The researchers evaluated the platform using human surgical glioma specimens. According to the methods section, human tissue samples were collected from 17 patients undergoing brain tumor resection, including astrocytoma, oligodendroglioma, glioblastoma, childhood diffuse hemispheric glioma, and low-grade glioma. Mouse brain tissue and other organ tissues were also used to validate tissue clearing, imaging depth, image restoration, and preservation of tissue architecture.
In glioma specimens, ULTRA revealed 3D pathological features that conventional two-dimensional sections may not fully capture. The platform visualized key histological features across different glioma subtypes, including nuclear atypia, microcystic changes, abnormal mitosis, microvascular proliferation, and necrotic regions. Rather than appearing only in isolated sections, these features could be observed continuously across three-dimensional tissue volumes.
Most importantly, ULTRA was applied to glioma infiltration margin analysis. In a glioblastoma margin specimen, the researchers processed and analyzed approximately 1 mm³ of tissue using ULTRA. The total workflow, including tissue clearing, SRS imaging, and AI-based virtual staining, took about 30 minutes. The team then used cellularity analysis and a 3D convolutional neural network to generate an ULTRAscore, representing the probability that a given 3D tissue block contained tumor. This enabled segmentation of the tissue into dense tumor, sparse infiltrative tumor, and non-tumor brain regions.
This 3D view revealed strong depth-dependent heterogeneity at glioma margins. Within the same tissue volume, some individual two-dimensional planes appeared nearly normal or diagnostically uncertain, while adjacent planes showed tumor cell infiltration. When independent pathologists reviewed only single 2D sections from the same volumetric dataset, they missed tumor in certain depth ranges where tumor was present in adjacent planes, with miss rates of 20% and 23%. In direct benchmarking, the 3D CNN also outperformed a 2D CNN, with an AUC of 0.965 compared with 0.909.
The study further showed that ULTRA could help address intraoperative uncertainty in MRI-defined edema regions and areas affected by navigation drift. In glioma surgery, FLAIR-hyperintense edema regions on MRI cannot reliably distinguish tumor infiltration from reactive or non-tumor brain tissue. In representative samples, ULTRA detected histological tumor involvement in radiographically ambiguous edema regions and in tissue areas that appeared normal on neuronavigation and were negative on neurophysiological monitoring. These findings suggest that ULTRA may serve as a specimen-based complement to imaging navigation and neurophysiological monitoring by providing additional histological context during surgery.
The findings should still be interpreted within clear limits. This study represents method development and validation using clinical specimens, not a large-scale clinical outcomes trial. ULTRA should not yet be described as a replacement for frozen section analysis, nor does the study show that the platform improves survival, recurrence rates, or postoperative neurological outcomes. The authors also note several technical limitations, including axial point-spread-function anisotropy, the relatively slow speed of the diffusion-based deep learning model, possible needs for tissue-specific optimization, and the current concentration of stimulated Raman scattering systems in research or specialized clinical centers. Further large-scale prospective clinical studies will be needed to evaluate its reliability, efficiency, and clinical value in real intraoperative workflows.
Overall, this study presents a new route for bringing 3D histology into the operating room. By combining rapid tissue clearing, label-free optical imaging, and AI-based virtual staining, ULTRA makes it possible to visualize glioma infiltration margins in three dimensions within an intraoperative timeframe. Its significance lies not only in making histology faster, but also in expanding intraoperative pathology from a small number of two-dimensional sections to richer volumetric tissue information, providing a new technical foundation for brain tumor margin assessment and surgical decision-making.
Article information:
Zhijie Liu, Yingying Li, Lingchao Chen, Minqian Wei, Mian Wei, Yuchen Sun, Tongqi Wang, Haixia Cheng, Xing Liu, Minbiao Ji, Lixue Shi.
Ultrarapid deep 3D histology enables intraoperative mapping of glioma infiltration . Cell , 2026.
Original article link: https://doi.org/10.1016/j.cell.2026.07.026
Cell
Computational simulation/modeling
People
Ultrarapid deep 3D histology enables intraoperative mapping of glioma infiltration
3-Aug-2026