Glioma surgery has long been a race against an invisible enemy. Tumor cells can spread microscopically through apparently healthy brain tissue, extending beyond the margins visible to surgeons, conventional imaging, or even standard intraoperative pathology. Removing too little tumor can leave behind infiltrating cancer cells, while removing too much can damage regions responsible for speech, movement, memory, or other essential functions. A new platform reported in Cell aims to give surgeons a more detailed view of this hidden tumor landscape by producing three-dimensional histology from brain tissue in approximately 30 minutes.
The technology, called ULTRA—short for ultrarapid cleared stimulated Raman with AI—combines rapid tissue clearing, stimulated Raman scattering microscopy, and artificial intelligence-based virtual staining. Developed by researchers from Fudan University, Huashan Hospital of Fudan University, Zhongshan Hospital of Fudan University, Beijing Neurosurgical Institute, Capital Medical University, and collaborating institutions, the system is designed to examine fresh or fixed surgical tissue without conventional staining or physical sectioning. Instead of reducing a specimen to a small number of thin slices, ULTRA preserves and analyzes tissue architecture across a three-dimensional volume, potentially revealing how tumor cells are distributed through depth.
The central challenge addressed by the platform is the mismatch between the three-dimensional nature of gliomas and the two-dimensional nature of routine intraoperative pathology. Frozen-section analysis can provide rapid information during an operation, but it samples only selected tissue planes and may be influenced by the location of the sample, section thickness, freezing artifacts, and staining quality. A tumor cell population located between sections can therefore be missed. Three-dimensional histology offers a more complete view, but conventional approaches often require lengthy fixation, chemical processing, fluorescent labeling, staining, and imaging—procedures that can take many hours or days. ULTRA is engineered to shorten this process to a timescale closer to surgical decision-making.
The first stage of the system uses a rapid clearing method developed to make brain tissue sufficiently transparent for deep optical imaging while remaining compatible with stimulated Raman scattering, or SRS, microscopy. SRS detects chemical-bond vibrations within molecules by using precisely tuned laser beams, allowing researchers to identify intrinsic signals from major tissue components such as lipids and proteins without adding fluorescent dyes. Because the method is label-free, it can preserve chemical information that might be altered or obscured by conventional staining. In ULTRA, SRS imaging captures volumetric signals from millimeter-scale tissue samples, providing a molecularly informative foundation for reconstructing cellular and tissue morphology.
Artificial intelligence then transforms the raw optical data into a more interpretable form. The system uses three sequential computational modules. A diffusion-model-based, depth-aware denoising network restores image quality as the SRS signal becomes weaker deeper inside the specimen. A conditional generative adversarial network predicts the protein channel from the lipid channel, reducing the need to acquire both channels throughout the entire volume and thereby lowering imaging time and hardware demands. Finally, a virtual hematoxylin-and-eosin, or H&E, staining model converts the SRS information into images that resemble the familiar appearance of standard pathology slides. The resulting virtual histology does not replace a pathologist’s judgment; it is intended to make complex three-dimensional optical data easier to inspect and interpret.
The researchers tested ULTRA on human surgical specimens from 17 patients undergoing brain tumor resection. The samples included astrocytoma, oligodendroglioma, glioblastoma, childhood diffuse hemispheric glioma, and low-grade glioma. Mouse brain tissue and samples from other organs were also used to assess the clearing procedure, imaging depth, image restoration, and preservation of tissue structure. In the glioma specimens, the platform visualized features associated with tumor biology, including nuclear atypia, microcystic changes, abnormal mitotic activity, microvascular proliferation, and necrosis. These characteristics could be followed across tissue volumes rather than viewed as isolated findings in separate two-dimensional sections.
The most significant demonstration involved the analysis of an infiltrative glioblastoma margin. The researchers processed approximately one cubic millimeter of tissue, completing clearing, SRS imaging, and AI-based virtual staining in about 30 minutes. They then combined cellularity measurements with a three-dimensional convolutional neural network to produce an “ULTRAscore,” a probability estimate for whether a particular three-dimensional tissue block contained tumor. This enabled the team to distinguish regions of dense tumor, sparse infiltrative tumor, and non-tumor brain tissue. The approach provided a depth-resolved map of tumor infiltration, showing how the apparent boundary changed from one plane to the next.
That depth-dependent variation exposed a weakness of relying on individual two-dimensional views. Some planes within the same specimen appeared nearly normal or diagnostically uncertain, while neighboring planes contained infiltrating tumor cells. When independent pathologists examined only selected two-dimensional sections from the volumetric data, tumor was missed in specific depth ranges; the reported miss rates were 20% and 23%. In a direct comparison of computational approaches, the three-dimensional neural network achieved an area under the receiver operating characteristic curve of 0.965, compared with 0.909 for a two-dimensional network. These findings suggest that incorporating spatial context may improve the ability to recognize subtle infiltration, although the results require confirmation in larger clinical studies.
ULTRA may also provide information in regions where surgical imaging is ambiguous. Glioma-associated edema that appears hyperintense on fluid-attenuated inversion recovery MRI can contain infiltrating tumor, reactive tissue, or non-tumor brain, and MRI alone cannot always distinguish among these possibilities. In representative specimens, ULTRA identified tumor involvement in radiographically uncertain edema regions and in areas that appeared normal on neuronavigation and were negative on neurophysiological monitoring. The technology could therefore complement, rather than replace, MRI, navigation, monitoring, and frozen-section pathology by supplying additional tissue-level evidence during an operation. However, the study was a technology-development and clinical-specimen validation project, not a trial showing improved survival, recurrence rates, or neurological outcomes. The authors identify remaining challenges, including anisotropy in the axial point-spread function, the relatively slow diffusion-model computation, the possibility that tissue-specific optimization will be necessary, and the limited availability of SRS systems in routine hospitals. ULTRA is not yet a substitute for established diagnostic methods, but it offers a new way to bring rapid, label-free, three-dimensional histology closer to the operating room and could eventually change how surgeons and pathologists define the margins of infiltrative brain tumors.
Subject of Research: People
Article Title: Ultrarapid deep 3D histology enables intraoperative mapping of glioma infiltration
News Publication Date: 3-Aug-2026
Web References: https://doi.org/10.1016/j.cell.2026.07.026
References: Zhijie Liu, Yingying Li, Lingchao Chen, Minqian Wei, Mian Wei, Yuchen Sun, Tongqi Wang, Haixia Cheng, Xing Liu, Minbiao Ji, and Lixue Shi. “Ultrarapid deep 3D histology enables intraoperative mapping of glioma infiltration.” Cell, 2026. DOI: 10.1016/j.cell.2026.07.026
Image Credits: Lixue Shi
Keywords: glioma, brain cancer, glioma infiltration, three-dimensional histology, stimulated Raman scattering microscopy, artificial intelligence, virtual staining, intraoperative pathology, tissue clearing, neurosurgery, tumor margins, cancer research
