A new artificial-intelligence system could help pathologists search through three-dimensional tissue specimens without forcing them to inspect every microscopic layer by hand. The framework, called TRICARE, is designed to identify the most medically suspicious two-dimensional cross sections hidden inside large 3D pathology datasets. Its developers say the approach could make emerging volumetric imaging technologies more practical in hospitals, where the amount of data produced by a single specimen can rapidly exceed the time available for human review. Rather than replacing pathologists, TRICARE is intended to act as a sophisticated triage assistant, directing attention toward tissue levels that are most likely to contain cancer or precancerous changes while preserving expert clinicians as the final decision-makers.
Conventional histopathology offers an extraordinarily detailed view of tissue, but it does so through a narrow window. A biopsy is typically sliced into thin sections, mounted on glass slides and stained so that cells and tissue architecture can be examined under a microscope. Although these images can reveal crucial diagnostic features, each section represents only a tiny fraction of the original specimen—often less than 1% of the total biopsy volume. A lesion may therefore be missed if it lies between sampled sections or if its most important features are not captured in the selected planes. This problem becomes especially significant when disease is spatially patchy, with abnormal regions distributed unevenly throughout a specimen.
Three-dimensional pathology seeks to overcome that sampling limitation by imaging tissue throughout its depth without destroying it. One of the technologies enabling this shift is open-top light-sheet microscopy, which illuminates tissue with a thin sheet of light and captures fluorescence or other optical signals across broad areas. After suitable preparation, large clinical specimens can be scanned in three dimensions at high resolution, producing a volumetric map in which cells, glands and other structures can be examined across many consecutive depths. The result is more comprehensive than a conventional slide series, but it also creates a formidable data-management problem: a single tissue volume may contain hundreds or thousands of potential viewing levels, each resembling a digital pathology slide.
TRICARE addresses this challenge by assigning a risk score to every two-dimensional level within a 3D tissue volume. The system is based on deep learning, a class of machine-learning methods that uses multiple layers of artificial neural networks to recognize patterns in complex data. In this case, the model does not evaluate each image as an isolated snapshot. Instead, it incorporates information from a selected group of neighbouring depth levels, allowing it to interpret local three-dimensional context. That distinction is technically important because many pathological structures extend across several planes. A suspicious gland, fragmented lesion or evolving tissue boundary may appear ambiguous in one section but become much clearer when adjacent levels are considered together.
The researchers compared this context-aware strategy with models that make predictions from individual 2D levels alone. According to the study, TRICARE performed better than approaches that ignored neighbouring sections. The advantage reflects a basic property of biological tissue: disease is not distributed as a collection of unrelated flat images. Tumours and precancerous changes have continuity, shape and spatial relationships that unfold through depth. By examining a limited neighbourhood around each level, the model can use that continuity as evidence, potentially reducing errors caused by folds, artifacts, staining variation or anatomically misleading views. The system can then rank the full stack of sections, placing the most concerning levels at the front of a pathologist’s review queue.
The team evaluated the framework in two clinically important settings. The first involved prostate cancer biopsies, where disease can be small, irregularly distributed and difficult to characterize from limited sampling. In that use case, TRICARE was used for risk stratification, separating tissue levels according to the likelihood that they contained high-risk pathological features. The second focused on endoscopic biopsies from patients with Barrett’s esophagus, a condition in which the lining of the esophagus changes and can progress to dysplasia or cancer. Detecting these changes is a central goal of surveillance, yet abnormal areas may be sparse and easily overlooked when tissue is assessed through conventional sampling. In both applications, the researchers report that AI-assisted review of 3D pathology showed potential to improve the identification of high-risk disease compared with standard slide-based workflows.
The proposed workflow is deliberately conservative. TRICARE does not issue a final diagnosis or remove the pathologist from the process. Instead, it acts as a filter for an enormous image set, highlighting the levels that deserve priority and allowing clinicians to examine those areas first. A pathologist could then review the flagged sections, inspect surrounding tissue, and make a diagnosis using clinical information and professional judgment. This design may be especially important for the introduction of AI into pathology, a field in which false negatives can have serious consequences and unexplained automated decisions can be difficult to accept. By retaining a human expert at the end of the chain, the technology offers a lower-risk route toward using 3D datasets in routine assessment.
The potential workload reduction is substantial. If a 3D scan contains hundreds of relevant levels, reviewing every one with equal attention could turn a single case into a lengthy, impractical task. A triage system can transform that problem into a ranked investigation, concentrating human effort on the most informative regions while still preserving access to the complete volume. This does not mean that unflagged areas would automatically be considered normal. Rather, the model could help clinicians decide where to begin and which parts of the specimen require especially careful evaluation. In a future clinical laboratory, such prioritization could make comprehensive imaging more compatible with the time pressures of diagnostic medicine.
Important challenges remain before systems such as TRICARE can become routine clinical tools. Deep-learning models must be tested across different hospitals, scanners, staining protocols, tissue-preparation methods and patient populations. Their performance must also be measured not only by technical accuracy but by whether they genuinely improve diagnostic sensitivity, reduce missed lesions and fit safely into existing laboratory workflows. Three-dimensional pathology itself requires specialized imaging, data storage and computational infrastructure, while pathologists will need interfaces that make volumetric information intuitive rather than overwhelming. Even so, the study points toward a new model of pathology in which diagnosis is not restricted to a handful of thin slices. By combining comprehensive tissue imaging with spatially aware AI triage, TRICARE suggests that the future of biopsy analysis may involve seeing more of the specimen while asking humans to spend their time where it matters most.
Subject of Research: Deep-learning triage of three-dimensional pathology datasets for efficient detection and assessment of high-risk tissue regions.
Article Title: Deep-learning triage of three-dimensional pathology datasets for comprehensive and efficient pathologist assessments.
Article References: Gao, G., Yan, R., Song, A.H. et al. “Deep-learning triage of three-dimensional pathology datasets for comprehensive and efficient pathologist assessments.” Nature Biomedical Engineering (2026). https://doi.org/10.1038/s41551-026-01760-1
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41551-026-01760-1
Keywords: 3D pathology, artificial intelligence, deep learning, digital pathology, light-sheet microscopy, cancer detection, prostate cancer, Barrett’s esophagus, dysplasia, pathology triage
Tags: 3D tissue imagingAI-assisted pathology triageAI-driven cancer diagnosticscancer detection in tissue specimenscomputational analysis of large pathology datasetsdeep learning in pathologydigital pathology and AI integrationhistopathology image analysis toolsmedical imaging data managementmicroscopy and histopathology advancementstissue specimen analysis automationvolumetric imaging technologies in medicine
