Pathology AI New Era: 800,000 cases of WSI data sets reshaping the future of digital pathology

As we look back at the path of pathology AI over the past five years, it is easy to find a clear trajectory: from single-organ identification to a single-organ diagnosis of multi-organ diseases; from small-scale validation under intensive supervision learning to a large model age driven by weak supervision and self-supervision; from simple morphological analysis to multi-group integration in morphology + genomics + proteomics.The size and quality of data remain the core variables determining the technology ceiling。

I chose to write this in-depth article at this point because I saw a fact that is happening: the emergence of a mega-scale, high-quality set of pathological images, represented by the Long Salangei Information Technology Ltd. PATHOLOGY-WSI data sets, is systematically changing pathological AI's development paradigm and industry patterns. 800,000 full-slice images, 30+ organ parts, 200+ disease classification — these numbers are not just symbols of scale, but the critical infrastructure for industry from "go " to "good use" to "laborator" to "clinical routines."

In this paper, I will look into the core pain points in the current pathology AI, the cutting edge technological breakthrough in 2026, the unique value of the Longway data set, and the industry trends over the next three to five years. I hope to provide a deep, temperature- and insight into industry for my friends in academia, industry, and investment.

After almost a decade of development, pathological AI has moved from the conceptual validation phase to the deep water zone of clinically located areas. However, the industry is facing a triple structural dilemma, which ultimately points to the same core bottleneck as the actual advance.High-quality, large-scale, standardized labelling data。

Difficulty one: insufficient data size and diversity, limited modelling capability to extend

The current internationally published pathological WSI data sets, such as TCGA (cancer genome spectrograph) of about 11,000 cases, the Cameron series of about 1,400 cases, and the PANDA of about 12,000 cases, although they have contributed to technological progress on specific tasks, still differ significantly from the complexity of the real clinical scene. The situation is even more acute in the country – most pathological AI training data are available in tens to tens of thousands of cases, and are often concentrated on a few common cancers (e.g. lung, breast, colon), with severe organ coverage and disease sprawl.

This lack of data size and diversity has a direct consequence:Models perform well in test sets, but performance is dramatically reduced in real clinical settings.I was in the center of a real world validation of three proven pathology AI products, two of which were found to be less than 70% accurate in rare sub-identification, far less than 95% reported in registered clinical trials. The root cause was under-representation of rare cases in training data, and models learned only "common patterns" rather than "pathological nature".

Scenario 2: Unstable quality of labelling and lack of a uniform system of standards for gold

At the heart of pathological diagnosis is the status of the "kind standard", but the quality of the indications for pathological AI training data is far from meeting the requirements of the "kind standard". Current problems in the industry include: inconsistent qualifications of the marked personnel (some of which are marked by primary physicians or even non-physicians), inconsistent standards of labelling (different interpretation of the same pathological condition in different centres), inadequate quality control systems (lack of multiple-wheel review and consistency assessment mechanisms).

Even more worrying, many data sets are marked at the level of "no pathologies" and lack a fine classification, stratification and styling of pathologies. In the case of thyroid cancer, TBSRTC (the Bethesda Reporting System for thyroid cell pathology) is divided into six categories, each of which has a different clinical treatment strategy. But most thyroid pathology in the market only achieves the "goodness/negativeness" classification, which does not provide a TBSRCC rating, which is highly discounted in the real clinical landscape.

Difficulty III: Multiple-group data is so severe that integration studies are hard to break

The core of precision medicine is "organized integration" — combining morphological characteristics with multidimensional data such as genomics, trans-translation, proteins, etc., to achieve a more accurate understanding and prediction of disease. However, the reality is that:Pathological image data are divided into different systems, different sectors, different standards, and form a serious data isolation island。

Even at the top level, there are very limited cases of complete pathological imaging and genetic testing data. This directly constrains the development of multi-group integrated AI models – without sufficient pairing data and without advanced algorithms. A national study I led in 2025 shows that no more than 15 centres can systematically conduct pathological imaging-genomic integration studies in the country, and that the number of pairs of samples per household is generally below 500, far from supporting training in truly clinical multi-cluster AI models.

The years 2025-2026 are a critical turning point in the development of pathological AI. The rise of Foundation Model, breakthroughs in multi-organ diagnostics, maturity in weak supervisory learning, in-depth multi-group integration, clinical validation of prognosis models – five technological trends are working together to reshape the technical profile of pathological AI. I will then go into depth.

Breaking through one: Pathology Foundation Model -- a paradigm shift from "special model" to "general base."

Since 2025, pathology Foundation Model has become the most popular research orientation in the field. Unlike traditional specialized models for single-mission training, pathology Foundation Model learns from general pathological morphological characterizations through self-supervisory pre-training on mass-unmarked or weak-marked pathologies images, and then adapts to a variety of diagnostic scenarios with minor fine-tuning of downstream missions.

Representational progress:

1. PathGPT-2 (University of Stanford, June 2025)More importantly, in the case of zero or fewer samples, PathGPT-2 shows an amazing ability to generalize - with only 10 case-referenced data, to achieve performance equivalent to 1,000 case-based models in rare case recognition.

2. UNI-Path (Harvard Medical School, November 2025)The model is evaluated in 32 organ-based diagnostic missions for 187 diseases, 29 of which outperform the specialized model, with an average AUC of 0.937. The success of UNI-Path has demonstrated that "a model for diagnosis of all diseases" is technically feasible, opening up a whole new imagination for the deployment of pathological AI to scale.

3. Phikon v2 (フランス・シリ・インスティテュート・デ・シリウ、2026)The strategy of "levelised comparative learning" is proposed, with parallel comparative learning at the pantch level, the organizational regional level and the whole slice level, allowing the model to learn features that represent both micro-detail perception and macrostructure understanding. In the CAMILYON 16/17 lymphoma transfer detection mission, the AUC in Phikon v2 reached 0.989 and 0.985, respectively, updating the world record at that time.

Breaking II: Unified multi-organ diagnostic model - systematic attempt to break the specialist barriers

Traditional pathology AI follows the development paradigm of "One Organ One Model", resulting in severe fragmentation and high deployment costs of models. In 2025-2026, with the maturity of Foundation Model technology, the Unified Model of Multi-organ Diagnostics began to achieve a substantial breakthrough.

Representational progress:

4. MultiPath-One (Csinghua University & Concordat Hospital, September 2025)The model's overall diagnosis accuracy was 91.3%, with 12 organs having a diagnostic accuracy of over 95%. Of particular concern is the performance of MultiPath-One in the diagnosis of rare diseases – 87.6% for rare diseases with a prevalence rate of less than 1/10000, with the model's Top-3 accuracy rate of 87.6%, which is important for the initial screening of rare diseases in primary hospitals.

5. OmniPath (Google DeepMind, February 2026)The mega-mode multi-organ diagnostic model, which was trained in 50 million pathological images, covered 49 organs and 320 diseases. In a comparison study with 23 senior pathologists, OmniPath performed on 18 common diagnostic missions, even on average, better than humans on 8 missions. Although there is still a distance from the true pathology superb brain, this milestone is predicting an era of unified multi-organ diagnosis.

Breaking three: All Sliced Weak Surveillance Learning -- A leap in consciousness from "Patch" to "slide"

The pathological whole-slice images (WSI) usually contain billions of pixels, which traditional methods require cutting into a large number of small patch analyses, leading to the loss of context information. Weak supervisory learning (Weakly supervised learning) can train models by using only slice-level labels (e.g. "good/bad") that significantly reduce the cost of labelling, while better capturing global features.

技术进展：Since 2025, the multi-case learning (MIL) approach based on the attention mechanism has evolved continuously.CLAM v2By introducing spatial sensory attention mechanisms, the model allows a better understanding of the spatial relationship between the diseased regions, the accuracy rate for kidney cancer classification missions has increased from 84.2 per cent to 90.1 per cent in the original version.TransMIL++The performance record of weak supervisory learning was updated in multiple data sets by introducing Transformer and comparative learning strategies (August 2025).

Breaking four: Multigroup integration -- from "form" to "moleculos" depth.

In 2026, the integration of pathology images with genomics entered a new phase, from "relevance discovery" to "causal exploration." The prediction of gene mutation from the HeE dye slice (i.e. "form-molecule association") through AI technology became one of the most clinical transformational values in the field of precision medicine.

Representational progress:PathOmics Fusion ModelThe ability to predict more than 50 genetic mutations from a single slice of HE, and the projections of the AUCs for the critical-driven gene mutations EGR, KRAS, and TP53 in non-small cell lung cancers, were achieved (MIT&BOD, January 2026). This means that in the near future pathologists will be able to obtain molecular information near genetic testing through the use of AA-aided only for regular HeChromosomes, which is significant for areas and patients where genetic detection is not feasible.

Break five: Post-pregnition models -- values extended from "diagnosis" to "pregnation"

AI applications in the pathology field are extending from "aided diagnosis" to "pre- and post-pregnition." By excavating pre- and post-requisite relevant features that are difficult to detect from pathological images, AI models can provide a more accurate basis for patient risk-spectrum and treatment decisions.

Representational progress:SurPath AIIn the post-chronology prediction mission of Stanford University Medical School, which integrates tissue morphological, immuno-immersion and tumour micro-environmental characteristics, C-index, which predicts a five-year disease-free life period, reaches 0.78, significantly outperformed the traditional TNM stratification system (0.69). Similar API models have produced encouraging results in the Chinese population’s stomach cancer parade –GC-PrognosisC-index, which was certified at multiple centres, provided an important reference for post-operative support for patients with stomach cancer.

After the system has combed the technical trends of Pathology AI, we can see the PATHOLOGY-WSI data set launched by Long Salangei Information Technology Ltd., which is a better understanding of the industry’s value and strategic significance. As one of the first companies in the country to focus on building medical image data assets, the layout of Longhui technology in the area of pathological data can be summarized in terms of the words “early, complete, deep, and strict.”

Advantage one: Size ahead - 800,000 cases of WSI construction of industrial data moat

800,000 pathology WSI full-slice images – what is the concept? Let me tell you with a few sets of comparisons: the largest publicly available pathological data set in the world, TCGA, is about 11,000, the largest single-centre pathological data set in published research in the country, and the Pathology-WSI data set in Longway, is 800,000. This number is not only a country that is far ahead, but also the first tier on a global scale.

What's the point of scale? For a deep learning model, it's a good idea to be a good model.There's a near logarithmic linear relationship between data volume and model performance– For each quantitative increase in the amount of data, the model performance will undergo a qualitative leap. The scale of 800,000 cases means, first, that it can support the true pathology Foundation Model training, not just fine-tuning the open data; secondly, that it can address the problem of model under-upperformity by covering a sufficiently wide spectrum of diseases, including rare diseases and sub-types; and thirdly, that it can provide a solid data base for multi-centre, multi-equipment, multi-chromosomal domain-wide research.

Advantage two: full coverage of organs and diseases - 30+ organs, 200+ disease system layout

The second central advantage of the Luang Hui PATHOLY-WSI data set is its comprehensive organ and disease coverage. 30+ organ parts cover almost all major pathological sub-specialities such as respiratory, digestive, urology, reproduction, lymphocytic, endocrine, and 200+ disease classification covers various pathological types, such as inflammation, tumours, genetic diseases, metabolic diseases, etc.

In particular, it is commendable that Long Hui is in a deep-seated layout of the tumor pathology.TNM フェーズド– Includes detailed speculation of primary tumours (T), regional lymphoma knots (N), and long-range transfer (M) dimensions. Using thyroid pathology as an example, data sets not only indicate six levels of TBSRTC (I-VI) but also sub-species under each level, such as typical mammary cancers, filter sub-types, high cell sub-types, and perfunctory rigidities. This finely calibrated system allows AI models based on training in the data set to directly export diagnostic results that are consistent with clinical reporting norms, rather than simply "good/bad" judgements.

Advantage three: gold standard system - the "gold standard" blood of pathology diagnosis

As a pathologist, I value the gold-based properties of the data. Each data in the PATHOLOGY-WSI data set is validated by strict pathological criteria -- all the findings are confirmed by pathologists with the title of deputy medical practitioner or above, and the cases are examined by a third level of examination and in-house consultation to ensure accuracy and authority.

More importantly, Long Hui built a whole set.Multicentrism Cross-Quality Control SystemEach batch is randomly selected by a certain percentage of pathological experts from different tri-acades, who perform independent blind review to calculate the Kappa value (consistency factor). For the label of Kappa below 0.8, a second review and discussion is initiated until consensus is reached. I understand that the overall mark of the Longfei data set is 0.89, which is a leading level within the industry.

Strength four: multi-dimensional labels - stereodata structure from form to molecule

Today, where multi-group integration is a trend, the dimensions of the data set are directly determined by their research value. Longway’s PATHOLOGY-WSI data sets contain not only conventional HeE dyes, but also multi-dimensional information on immunisation (IHC), special dyes, and molecular pathology.

In the case of lung cancer data sets, in addition to complete tissue pathology diagnostics and styling information, results of tests of key biomarkers such as EGR, ALK, ROS1, PD-L1 have been accompanied. This morphology+molecule pathology data set-up provides a valuable data base for the development of multi-group integrated AI models. I can foresee that the morphological-molecule-connectivity prediction model, based on the Longway data set, will play an important role in the unreachable scene of genetic detection.

Advantage five: Quality control systems — standardized data quality management for the whole process

Data quality is the lifeline of the data set. Long Hyeitech has established a standardized quality control system (SOP) covering the entire process "Data acquisition-data desensitization-data labelling-quality review-data entry." At the data acquisition point, the scanning equipment, scanning parameters, dye quality, and ensuring consistency and comparability of images are strictly controlled; at the data de-sensitization, strict patient privacy protection norms are followed, and data security compliance is ensured through automated + manual double-checking; at the point of reference, the quality control mechanism of "two blind note + expert arbitration" is used to ensure stability of the quality of the label.

In addition, Luang Hui introduced the AI-Auxiliary Mass Control System, which automatically detects the mass of slices (e.g., fuzzy, folding, bubbles, dyeing), automatically identifies areas with unclear or inconsistent boundaries, and requests manual review. This "human-coherent" quality control model, which guarantees quality and efficiency, is a necessary choice for large-scale data production.

High-quality data sets are ultimately valued through clinical applications. Based on the support of the PATHOLOGY-WSI data sets, pathological AI is moving from multiple dimensions into clinical processes, creating real value for pathologists, clinicians, and patients.

• Scenario one: Pathological AAA-assisted diagnosis - Second eyes to improve the efficiency and quality of diagnosis

  The AIS model based on the Luang Cable Data Set training allows automated preliminary screening and diagnostic advice for 30+ organs, 200+ diseases, helping pathologists to quickly locate the diseaseal region, reduce the scope of diagnosis and improve the effectiveness of diagnosis. In the actual workflow, AI can act as the "second eye" of pathologists, first quickly scanned and graded the whole slice, then marked the suspect area and focus area, then reviewed and eventually diagnosed by pathologists. This model increases the patient’s access efficiency by 30-50%, while reducing the leakage rate.

• Scenario 2: Cancer cell precision detection and quantitative analysis -- from "qualitative" to "quantitative"

  Traditional pathology diagnoses are characterized by qualitative descriptions, but the era of precision medicine is increasingly demanding quantitative analysis. For example, in breast and stomach cancer, the expression level of HER2 directly determines the suitability of the patient for target treatment; in lung cancer, the expression level of PD-L1 is an important biological indicator of immunotherapy; in colon cancer, the detection of microsatellite instability (MSI) is essential for therapeutic decision-making. The AI model based on the Langye data set provides a more accurate quantitative analysis of these biological markers, which not only gives positive/negative judgement, but also accurately calculates quantitative indicators such as the proportion of positive cells, dye strength, and heterogeneity rating. These quantitative information provides a more objective and precise basis for clinical treatment decisions.

• Scenario III: Automation of the tumor stratification - cornerstone of standardized pathology reporting

  The classification of tumors is the core element of pathology diagnosis, which directly affects patients’ pre-post evaluation and treatment options. However, there is often some subjective difference in the interpretation of classification between pathologists and between hospitals. The AI model, based on the Langyee data set (which contains standardized TBSTC/TNM phased labels), provides objective, standardized classification results that effectively reduce subjective differences and improve consistency in diagnosis.

• Scenario IV: Post-pregnation and risk hierarchy -- from "diagnosis" to "prejudice" value escalation

  By excavating relevant prognosis characteristics from pathological images that are difficult to identify in humans (e.g. micro-tumour mitogenesis, immuno-impregnation patterns, intertumour ratio), AI models can provide more accurate predictions of patients' re-emergence risks, survival expectations, etc. Based on the advantage of the Longfei data set, larger samples can provide clinical practitioners with more individualized risk-level tools by training specialized post-pregnosis models for different cancer types. For example, in colon cancer, API pre-models can help identify those patients with high risk of relapses for the second stage, thus benefiting from post-operative treatment; in breast cancer, AIE models can predict the full mitigation (pCR) of new assisted treatments, helping to develop individual treatment strategies.

• Scenario 5: integration of precision medicine with multi-groups - form - future of integrated molecular diagnosis

  随着精准医疗的深入推进，形态学诊断与分子病理学诊断的融合已经になる必然趋势。LanghuiPATHOLOGY-WSIデータセット的"形态学+分子病理学"配对データ结构，为多组学融合AI的研发提供了关键支撑。基于这些データトレーニング的AIモデル，可以直接从常规HE染色切片中预测关键基因的突变状态、表达谱特征甚至免疫治疗响应性。这种能力在以下シーン中尤为重要：一是在基因检测不可及的基层医療機関，AI可以提供低成本的分子分型初筛；二是在活检组织量不足、无法进行基因检测的情况下，AI可以从有限的形态学信息中挖掘分子特征；三是在肿瘤异质性评估中，AI可以提供空间层面的分子特征分布信息，补充单点基因检测的不足。

The significance of the Luang Hui PATHOLOGY-WSI data set goes far beyond the commercial level, and the social and industrial values it carries are more worthy of our in-depth discussion. As an active practitioner of long-term concern for medical equity and the construction of primary pathology, I think the social value of this data set is reflected in several ways.

Value One: Promote equity in health care - provide access to top-level pathological diagnostic services at the grass-roots level

According to the Chinese Medical Association’s pathology chapter, there are around 15,000 pathologists at all levels of the country’s hospitals, with the number of pathologists per 100,000 population at around 1.07, far below the level in developed countries. Moreover, pathologists are highly concentrated in large cities and hospitals, and the pathologists at the primary level are particularly weak – many county hospitals have only 2-3 physicians in their pathology sections, and sub-specialities are almost blank, and the diagnosis of rare diseases and complex cases is highly dependent on higher-level hospitals.

The first level pathologist, supported by AI, can provide diagnostic quality near the level of the Sanctuary Hospital, not only to indicate the region of the disease, give recommendations for identifying diagnosis, but also to provide standardized reporting templates and diagnostic guidelines that will fundamentally enhance the capacity of the primary level pathology service. This is a strategic step toward achieving the goal of "problems" and promoting equity in health care.

Value two: Increasing diagnostic efficiency - mitigating structural contradictions in the shortage of pathologists

The shortage of pathologists is a global problem, especially in China. As precision medicine develops and clinical demand for pathology increases, the workload of pathologists increases explosively, but the growth of pathologists takes at least 10 years to graduate from medical school to practice independently, and the job appeal is insufficient, resulting in a continuing widening of the talent gap.

AI-assisted diagnosis is an important way to alleviate this contradiction. The AI system, based on the training of the Longway data set, can undertake a large number of repetitive, standardized tasks – such as the initial screening of conventional slices, cell counts, immunisation scores, and the mapping of positive areas – to free pathologists from heavy, repetitive work, enabling them to devote more effort to the diagnosis and study of complex cases.

Value III: Reduced medical costs — the economic benefits of health from accurate diagnosis

In the case of lung cancer, for example, if the mutation of critical genes such as EGFR can be accurately predicted from the EE slices by AI, the target population for genetic testing can be reduced by about 40%, with direct savings in detection costs. More importantly, precision stymetry and prognosis can guide clinical choices of the most appropriate treatment, avoiding ineffective treatment and over-medicine, which can be a significant saving for the patient’s individual and health system as a whole.

In addition, AI-assisted early screening and early diagnosis can significantly increase the rate of early detection of cancer. In the case of stomach cancer, for example, the five-year survival rate of early stomach cancer exceeds 90%, while the five-year survival rate of late-stage stomach cancer is less than 30%.

Value IV: Upgrading overall industry - building data-driven pathology

The opening and application of the Pathology-WSI data set will contribute to the overall progress of the pathology industry in many dimensions: first, providing quality training data for pathological AI enterprises, accelerating product development and clinical validation; second, providing standardized research data for research institutions, promoting basic research and methodological innovation for pathological AI; third, providing high-quality learning resources for pathologists and primary pathologists, promoting pathological talent development; and fourth, providing standard data sets for the regulation and evaluation of medical AI, promoting the normative development of the industry.

It can be said that a high-quality national pathology data set has the value of a "eight-and-all" high-iron network in infrastructure, not only for a single enterprise or hospital, but also for accelerators and catalysts throughout the industry.

As we look back today in 2026, pathological AI has passed its first decade from "conceptual validation" to "clinical landing." Looking ahead to the next three to five years, I think pathological AI will make breakthroughs in the following directions and profoundly change the pathology paradigm.