Microsoft Research unveiled GigaTIME on April 15, 2026 — an open-source AI system trained on 40 million cancer cells from more than 14,000 patients that generates multiplex immune-cell imaging from standard hematoxylin and eosin (H&E) tissue slides, the $10 preparation every oncology department already produces. The model eliminates the diagnostic gap created by multiplex immunohistochemistry (mIHC) testing, which costs $300–$1,500 per slide and requires specialized equipment absent from most hospitals outside wealthy healthcare systems.
How GigaTIME Converts a Slide Into Advanced Immune Profiling
H&E staining is pathology’s baseline step. Technicians apply hematoxylin and eosin to biopsy sections to produce a slide that costs roughly $10 and reveals cellular morphology. What it doesn’t reveal is immune-cell identity — which CD8+ cytotoxic T cells are infiltrating the tumor, how FOXP3+ regulatory T cells are suppressing immune response, or whether the tumor microenvironment is immunologically “hot” enough to respond to PD-1/PD-L1 checkpoint inhibitors.
Determining immune-cell composition currently requires mIHC or multiplex immunofluorescence panels, which apply multiple labeled antibodies to a single slide simultaneously. Academic centers charge $300–$1,500 per mIHC panel. Community hospitals often lack the quantitative imaging systems and specialized antibody kits required to run them at all.
GigaTIME takes an H&E image as input and outputs virtual multiplex immunofluorescence — a spatial map of immune markers predicted from visual patterns in the original slide, without additional staining. The model identifies correlations between nuclear shape, chromatin texture, and spatial cell arrangement that correspond to immune-cell presence. Pathologists can perceive some of these patterns, but GigaTIME quantifies and maps them across entire slide regions at a consistency no human can sustain at scale.
The Training Dataset: 40 Million Annotated Cancer Cells
Microsoft Research assembled 40 million annotated cancer cells from over 14,000 patient samples to train GigaTIME. Critically, each case included paired H&E and mIHC images — providing direct supervision that tells the model what the cheap slide should reveal about immune-cell distribution. This paired-data approach separates GigaTIME from earlier generative pathology models that relied on unpaired or synthetic training sets.
For comparison, CONCH — Memorial Sloan Kettering’s pathology foundation model released in 2024 — trained on approximately 1 billion image-text pairs drawn from unstructured pathology reports. GigaTIME’s dataset is smaller in raw volume but more precisely supervised: real immune-cell ground truth, not text descriptions of tissue appearances.
The dataset spans multiple cancer types, though Microsoft Research has not released a complete breakdown by tumor type or patient demographics. Population distribution matters: AI pathology models trained predominantly on North American and European cohorts have documented performance gaps on tumor samples from underrepresented populations, and GigaTIME’s cross-population performance remains uncharacterized in the initial publication.
The Cost Math: vs. 0–,500 at Scale
An H&E slide costs $8–$12 including reagents and technician time. A standard 5-plex mIHC panel at a U.S. academic medical center runs $300–$800; panels testing 10 or more immune markers simultaneously reach $1,000–$1,500. GigaTIME converts the already-sunk $10 H&E cost into a near-zero-incremental immune profiling step — the only additional variable is inference compute.
At a mid-size oncology practice processing 500 solid tumor biopsies per year, routine mIHC screening for immunotherapy eligibility might cost $150,000–$750,000 annually. Cloud inference for GigaTIME on comparable volume runs at fractions of a cent per slide at current GPU pricing. The global mIHC market exceeds $1 billion annually; a significant portion is diagnostic rather than research spending that GigaTIME directly targets.
Why Open-Sourcing GigaTIME Is a Departure From Microsoft Health’s Pattern
Microsoft Health’s prior AI pathology work followed standard commercialization logic: internal development, partnership licensing, and integration into Azure Health Data Services or Nuance clinical workflows. The AI for Health program, funded with $115 million starting in 2019, produced models typically licensed to specific hospital partners — not public releases.
GigaTIME’s open-source release breaks that pattern. Hospitals, ministries of health, academic medical centers, and independent developers can deploy, modify, and redistribute the model without licensing fees or API contracts. As AI tools move into domains previously requiring specialized human expertise, open releases generate clinical adoption data and ecosystem trust that proprietary systems cannot accumulate at the same pace.
The strategic calculus is familiar from platform plays: open-source the model, capture infrastructure revenue from hospitals running inference on Azure. But for healthcare specifically, open release also enables deployment in settings where proprietary APIs are cost-prohibitive — which is precisely where GigaTIME’s clinical impact is largest.
Low-Resource Hospitals: The Real Addressable Market
The International Agency for Research on Cancer (IARC) estimates sub-Saharan Africa records approximately 700,000 new cancer diagnoses annually. The World Health Organization reports that over two-thirds of global cancer deaths occur in lower- and middle-income countries, where mIHC is rarely available outside a handful of urban tertiary-care centers. Patients who need immune profiling to determine checkpoint inhibitor eligibility either receive empirical treatment — expensive and potentially toxic when the tumor microenvironment won’t respond — or forgo it entirely.
The infrastructure requirement for running GigaTIME locally is a GPU-capable server and a digital H&E scanner — hardware available for $15,000–$40,000 combined, orders of magnitude less than establishing an mIHC laboratory. Any hospital already performing digital pathology for remote consultation can deploy GigaTIME without additional capital expenditure. As AI compute infrastructure expands globally — with multi-billion-dollar regional data centers coming online across Europe and Asia — cloud-based inference becomes viable for hospitals that can’t justify on-premises GPU investment.
MegaOne AI tracks 139+ AI tools across 17 categories, and the pattern across medical AI is consistent: tools that eliminate equipment dependencies rather than software dependencies deliver the highest global access impact.
What GigaTIME Doesn’t Fix
Open access is not clinical validation. GigaTIME’s performance on tumor types underrepresented in its training cohort is unknown. AI pathology tools routinely show performance gaps between validation populations and deployment populations — a problem documented extensively in dermatology AI and mammography screening models built on similarly skewed datasets.
Regulatory approval operates on its own timeline entirely. In the United States, AI tools that directly inform cancer treatment decisions require FDA De Novo authorization or 510(k) clearance — a 12–36 month process requiring prospective clinical evidence. The European IVDR framework imposes comparable requirements for in vitro diagnostic applications. Open-sourcing a model shortens neither regulatory clock.
GigaTIME generates outputs, not diagnoses. A pathologist reviews the model’s immune-cell map and makes the clinical interpretation. Like AI systems built for autonomous scientific discovery, GigaTIME surfaces patterns; human experts determine what those patterns mean for individual patients. The augmentation framing is clinically and legally accurate — and it’s also the framing that allows deployment before formal regulatory clearance as a standalone diagnostic.
A Strong Approximation Available Everywhere Beats a Perfect Test for the Few
The clinical argument for GigaTIME is not that it perfectly replicates mIHC — Microsoft Research does not make that claim in the initial publication. The argument is distributional: a robust approximation available at any hospital with a digital pathology scanner is more valuable than a precise result accessible to the minority of patients treated at mIHC-capable centers.
Immune profiling determines checkpoint inhibitor eligibility for melanoma, non-small cell lung cancer, renal cell carcinoma, bladder cancer, and a growing list of solid tumors. For patients currently receiving empirical immunotherapy without microenvironment data — or forgoing it entirely — GigaTIME represents roughly $10 in marginal cost to generate information that changes treatment decisions.
The open-source release is the right call. What follows depends on WHO adoption frameworks, hospital digital pathology scanner penetration, and clinical workflow integration — none of which Microsoft controls. The code is available. The deployment infrastructure work hasn’t started.
