2026 Nature level breakthroughs with 5 million slices of data base from mSTAR to TITAN: pathology WSI base model

Pathological diagnosis is the "kind standard" for cancer diagnosis, and the maturity of the whole-slice digital imaging (WSI) technology has made it possible to digitize pathologies. However, the WSI gigapixel resolution and complex organizational structure make the traditional CNN model difficult to handle effectively. In 2025-2026, pathological WSI AI entered the "basic model" era - mSTAR (Nature Communications 2025) and TITAN (Nature Medicine 2025) through multi-modular self-monitoring learning, and performed excellently in 97 benchmark missions, marking the leap of pathological AI from "one mission model" to "general basic model".

Long Salang Langhui Information Technology Ltd. supports this technological breakthrough with a 5 million multi-organ pathology WSI data set. This data set is stored in the DITC VL Hole Slide Microscopy Image standard, covering HeE dye, IHC mono-label/multiplatform and special dye, multi-organ spectral coverage, and is an important data resource for training in basic pathology models.

The underlying modelling of pathology AI is far-reaching. In traditional models, each cancer type requires separate training models — lung cancer models, breast cancer models, stomach cancer models — which leads to the development of models that are costly and difficult to generalize. Basic models can significantly reduce the costs of pathological AI development and deployment by pre-training in a wide variety of data.

In 2025-2026, a "basic model competition" was launched in the pathology WSI AI field, with several top research teams and companies publishing their own pathology foundation models.

mSTAR (Multimodal Self-Taught Pretraining) published by Nature Commissions in December 2025, integrates three-modular data on WSI, pathological reports and genetic expression, using 26,169 slide-level model pairs, 32 cancer types, and over 116 million pattages. The performance in the cross-97 oncology baseline missions is excellent, ranking first in the 18 missions of the 21 pathological diagnostic data sets, and macroAUC has on average increased by 1.37 per cent.

TITAN, published by Nature Medicine in 2025, is an important indicator of the evolution of pathological AI from "aided diagnosis" to "autonomous diagnosis" through the visual supervision of self-learning and visual-linguistic alignment of 335,645 WSIs, not only for pathological diagnosis, but also for the production of pathological reports.

In February 2026, Nature published the CYBO-Clinical Autocell pathology Platform, which combines real-time optical whole-slice dialysis and marginal calculations to achieve cell-level classification and group morphological analysis. This is a milestone for Pathology AI to move from "research validation" to "clinical autonomy".

In addition, the HSFLA model (npj Digital Medicine 2026) achieved an accuracy rate of 95.6 per cent (by hand) on 1161 WSIs, and the automatic marking of the immersion area was consistent with artificial pixels by 86.6 per cent. The Prov-GigaPath series (Nature 2024, arXiv 2026) introduced high-impact reasoning-oriented variants such as GigaPath-Flash and Gigatime.

In terms of market size, the global digital pathology market is expected to exceed $1.5 billion in 2026, with AI-assisted diagnosis being the fastest growing area. China’s shortage of pathologists (3-4 pathologists per million population, well below 8-10 in developed countries) is acute, and the need for pathologist AI is particularly acute.

The core breakthrough of the pathology WSI AI 2025-2026 is in four ways.

The first is the enhancement of knowledge of multiple mosaics. The innovation of mSTAR is the integration of three mosaics of WSI images, pathological report text and genetic expression data. This multimodular convergence allows models to read not only pathological images, but also pathological reports and understanding genetic information, and achieve a multi-dimensional synthesis of pathological diagnosis. Eighteen of the 21 benchmark tasks rank first, proving the superiority of multi-mode learning.

The second is self-supervised/self-teaching. TITAN is self-supervised through 335,645 WSIs, and can learn the universal character of pathological images without manual labelling. This breakthrough has significantly reduced reliance on labeled data, making it possible to use large amounts of unmarked pathological data.

Third is the ability to produce reports. TITAN not only provides pathological diagnosis, but also produces structural pathological reports. This means that AI can automatically generate preliminary reports from the "Classer" to "Physician Assistant" for examination and revision by pathologists, thereby significantly improving the effectiveness of diagnosis.

The fourth is the clinical autonomy platform. The CYBO platform, published by Nature in 2026, has achieved the "sampling in and out of" autonomous cell pathology process by calculating the whole spectrum of the whole spectrum of the optical in real time, from the whole spectrum of the optic to the edge. It's a qualitative change in pathology from "assist" to "autonomy."

5 million WSI data sets in Longway Technology provide a key support for these frontier studies. The data sets cover multi-organ spectral systems, including various types of HeE, IHC and special dyes, and retain the original Level-0 data and all pyramid levels. Each slice is linked to a pathological diagnostic report and supports multi-level labels of slice, case, ROI and cell levels.

The Longhui Tech Multiorgan Pathology WSI image data set, with a size of 5 million slices, is an important data resource to support training in basic pathology models.

In terms of data collection, all slices are derived from the Unitary Pathology Section of the Sanchae Hospital, giving priority to the DITC VL Hole Slide Microscopy Image standard, while preserving original and OME-TIFF/BigTIFF 6.3.1 interoperability. All slices are subject to a 20-fold + scanning multiplier, retaining the Level-0 raw data and all pyramid levels, focal plane/optical channels and metadata.

For dye types, the main substance is HE dye; IHC mono/multi-label antibody layer; special dye by type; frozen/paraffin layer. This layer design enables models to learn organizational characteristics of different dyes and supports multi-chromosomal joint learning.

In terms of coverage, the multi-organ spectral system is covered, avoiding overconcentration by single cancers or single dye sources.

In terms of the system of signs, each slice is linked to the pathological diagnosis report, and includes TNM stages, immunisation results. Support is given to multi-level labels of the slice, case, ROI and cell levels, which are not to be used at different levels.

In terms of quality control, the three-layer process of control using the first standard of pathologist + deputy director and above experts to review the +conformity assessment is used.

In terms of compliance, all slices are legally authorized from a source and PHI fields are fully dissensitized.

The future development of pathology WSI AI will follow three paths.

The first is the continuous expansion of the base model. The mSTAR currently uses 26,169 model pairs and 116 million patch, but this is still small compared to the tens of millions of pathological slices produced globally each year. Larger and more diversified pre-training data will further improve the performance of the base model.

The second is adaptation of the Chinese population. Unlike Europe and the United States, where common cancers are spread (e.g. liver, stomach, and oesophagus cancer are more prevalent), models based on data training in Europe and the United States may be underperforming these cancers.

The third is clinical autonomy. From "aided diagnosis" to "automated diagnosis" is the ultimate goal of pathology AI. The autonomous cytology process of the CYBO platform (Nature 2026) predicts this direction. The data set of Longhui technology is linked by multilevel labelling and pathological reporting, which provides structured labels for training in autonomous diagnostic models.

The shift of pathology WSI AI from a "single-task CNN" to a "multi-modular base model" has revealed the decisive role of large-scale diversity data in the basic model age. The success of mSTAR and TITAN has demonstrated that data size, model diversity and marking depth are the three pillars of underlying model performance.

Long Shara'i Information Technology Ltd, which is based on more than 5 million organ pathology WSI data sets, is becoming an important data service provider for the pathology base model age. At a critical time when Nature and Nature Medicine levels have broken through the re-engineering of pathological AI, Long Slu's technology will continue to drive the autonomous release of pathological diagnostics with high-quality data.

The pathology AI in 2026 ushered in a fundamental shift from mission-specific models to basic model models. Basic models such as mSTAR and TITAN demonstrated that large-scale pre-training can be learned from generic pathologies, significantly reducing data labelling costs and modelling development thresholds by fine-tuning downstream with multiple diagnostic missions. This shift introduced new data requirements — beyond single-disease, single-tape-marked data sets — but required extensive, multi-pathological, multi-centre-based full-slice image pre-training language.

In terms of industry competition, the competition for the pathological AI basic model is essentially a competition for data size. International giants such as Paiga and PathAI have accumulated millions of pre-training data on WSI, while China is still in the initial stages of this area.