2026 Skin AI Age: industry map of 1 million SDS

Expert guide: Skin AI - Medically artificial intelligence 's "Stever"

I have experienced the technical integrations of skin disease diagnosis from "the eye plus experience" to "the skin lens plus pathology" in over 30 years of clinical and scientific work in dermatology. And the most exciting change in the last five years has been the rapid advance of artificial intelligence in the field of skin science.

Skin disease is one of the first areas of medical AI's breakthroughs – in 2017, a study by Stanford University team in Nature showed that deep learning algorithms have reached the level of dermatology specialists in skin cancer diagnostics. Since then, skin AI has been dubbed "the medically artificial intelligence pioneer." Yet, eight years later, despite the fact that the papers are full of papers, they are rarely actually clinically produced.

The core bottleneck, I think, remains data. Skin diseases range from 2,000 to a wide variety of clinical performances, with wide variations in skin damage from one race to another, from one part to another. To train a skin AI system that really has clinical practical value, it takes not tens of thousands and hundreds of thousands of people, but millions of high-quality markers.

In this article, I will look at the core pain points in skin AI from the perspective of a dermatologist, the breakthrough in front-line technology in 2026, and the strategic value of high-quality skin lens data sets for the industry as a whole. Skin AI spring may have arrived earlier than we thought.

II. Trade pain: The Four Gaps of Skin AI

2.1 Pathological divide: 2,000 skin diseases and "long end"

Skin pathology is an extremely complex discipline – more than 2,000 skin diseases are known, of which about 200 are common and more than 1800 rare. In clinical practice, 80% of patients suffer from 20% of common diseases, while the remaining 80% of rare diseases, although with low single disease incidence, together account for a significant proportion of dermatological clinics.

Most skin AI products can identify only a dozen or more common skin diseases, and are completely powerless at rare diseases. And what really tests doctor level in clinical practice is rare, unusual and unusual diseases. If AI only identifies common diseases, its value is reduced -- after all, general practitioners can diagnose common diseases, and patients come to the dermatology to understand the problems of "fail and incomprehensible."

2.2 Digital divide: "triple deficit" in open data sets

The most well-known internationally available skin lens data sets are the ISIC (International Organization for the Cooperation in Dermal Images), the latest version of which contains about 60,000 images. The country’s open data sets are smaller in size, with more than 130,000.

2.3 Panorama: the real world and the gap between laboratories

A common phenomenon in the skin AI field is that models that perform excellently in open data sets are "controversial" when they are in a real clinical setting. This is because the data sources of open datasets are relatively homogeneous, the equipment type, the conditions for filming, the composition of the population are homogenous, and the data in the real world are extremely different in quality - different brand skin mirrors, different angles and light, different colours and age groups, skin damage in different parts of the world.

I've been involved in a multi-centre validation of a skin AA product, and the results are thought-provoking: in R & D units, the model AUC reaches 0.95; in another hospital, the data dropped directly to 0.78. This difference of almost 20 percentage points is called the "wide-up divide." The only way to bridge this divide is to train models with more diverse, representative, large-scale data.

2.4 The trust gap: the "black box" dilemma of AI

Another deep challenge for skin AI is the question of "explainability". Most of the current AI models give only one diagnosis and probability value, but they cannot explain "why this diagnosis" — what are the characteristics of skin damage? These characteristics point to what disease, and what is the diagnosis?

For dermatologists, the diagnostic process is a process of reasoning based on skin-dermal characteristics. If AI gives only one answer and not one, it is difficult for the doctor to build trust in AI. That is why many skin-AI products are stuck in the top of the shelf after clinical trials.

III. 2026 Frontline Breakout: Five technological waves of skin AI

In 2025-2026, a new technological explosion occurred in the skin AI field. The integration of front-line technologies such as large models, polymodulates, Foundation Models, and dermal pathology has led to a series of exciting research results.

3.1 Large multi-modular model: from "seeing the disease" to "complex diagnosis"

The year 2026 is the year of the Large Skin Ai model. Traditional Skin Ai is classified on the basis of single skin mirror images, while a new generation of large polymodular models can process multiple mosaics of skin mirror images, clinical photographs, patient history descriptions, laboratory results, etc., and give more comprehensive diagnostic advice.

The Google DeepMind model, published in cooperation with King’s College London, uses a large visual-linguistic model structure, with a Top-1 accuracy rate of 83.6 per cent and Top-3 accuracy rate of 94.2 per cent for diagnostic missions that include 200 skin diseases, exceeding the performance of a median skin surgeon (Top-1 accuracy rate of 75.3 per cent). More importantly, DermGPT can produce diagnostic reports in natural languages, including skin-dermal description, identification of diagnostics, further examination recommendations, etc., that actually have the prototype of “AI dematology”.

At the national level, the team of the Sun-Zhi Hospital, affiliated with the Shanghai University of Transport Medical School, based on the SkinLM model developed by the Lang Hui DeRM-SKIN data set, has an accuracy of 81.3 per cent for the classification of 150 skin diseases and produces a better quality than a large model abroad for skin disease descriptions in Chinese and is more suitable for clinical scenes in China.

3.2 Foundation Model dermal preparation: building a "basic model" for skin sciences

Foundation Model is reshaping the whole AI field, including skin AI. At the end of 2025, Meta AI released DermFound - the first basic model of skin-oriented mirror images, which acquired a strong skin visual expression capability through self-monitoring pre-training on 10 million skin images (skin-bearing mirrors and clinical photographs).

On downstream missions, DermFund shows amazing performance: in blackoma detection missions, only 100 label images are fine-tuned, and AUC reaches 0.923; in 50 skin disease classification missions, the accuracy of the Few-Shot scene is 25 percentage points higher than traditional training from zero. This means that the base model can significantly reduce skin AI’s reliance on labeled data and accelerate AI’s landing in more diseases and scenes.

Of concern is the publication in March 2026 of the Chinese skin base model SkinFoundation-1M based on pre-training in the DERM-SKIN data set at the top dermatological hospitals of the United Nations in Longwaytech, which updated performance records on several Chinese skin AI missions and filled gaps in the domestic skin base model.

3.3 Leather description generation: leaps from "classification" to "description"

The output of traditional skin AI is "What is this disease?" The output of a new generation of AI is "What is this skin? " , which automatically produces structured skin-depletion descriptions. Not only does it greatly increase the interpretability of AI, but it also provides strong support for clinical case writing, teleconferences, teaching and other scenarios.

The DermCap model developed by the Stanford University School of Medicine, using an image-text comparison learning framework, automatically produces a description of the skin loss that meets the dermal norms of specialization, including: the type of skin damage (plaza, rubella, clots, nostrils, etc.), colour, size, shape, boundary, surface characteristics, distribution pattern, etc. In the doctor’s assessment, the description generated by DermCap is consistent with that of the senior dermatologist, well above the level of the senior medical practitioner (63%).

The clinical value of the technique of skin loss profiling is significant: it can assist low-age physicians in regulating the writing of medical records and improve their quality; it can provide standardized information on skin damage for remote consultations and reduce errors due to inconsistent descriptions; and it can also be used for automatic generation of skin lens reports, significantly improving the effectiveness of the treatment.

3.4 Rare skin disease recognition: the problem of long tail distribution

The first identification of rare skin diseases has been a technical challenge – too few samples and poor modelling. In 2026, the solution to this challenge was a major breakthrough. The RareDerm framework, developed by the Zurich University Hospital team, combined basic model migration, meta-learning and knowledge mapping techniques, achieved significant results in rare disease identification missions with only dozens of samples.

Specific data show that of the 50 rare skin diseases (one training sample of only 5-20) identified, the Top-1 accuracy rate for RareDerm is 62.7%, compared to the traditional method of 28.4%, which increases by 34 percentage points. While this figure is not perfect, given that many rare diseases are not necessarily diagnosed accurately even by dermatologists, the accuracy rate of 62.7% already has important clinical reference value - at least it can be shown that doctors consider the possibility of rare diseases and reduce the risk of leakage.

I think that the meaning of the AIS recognition for rare skin diseases is more of a "call" than a "confirmation" -- that AI can broaden the diagnosis of doctors by adding rare diseases to the diagnostic checklist when doctors do not think of rare diseases, thereby reducing the rate of leakage. This is as valuable as raising the overall level of treatment, not less than the accurate identification of common diseases.

3.5 Global crowdization: crossing borders of colour and ethnic origin

Skin AI has a problem with a disease - the performance of dark skin populations is significantly worse than that of light skin groups. This is because training data are mainly from white people, and dark skin samples are severely inadequate. In 2026, this "color bias" problem began to be systematically addressed.

The WSID data set, published by the World Skin Imaging Corporation, incorporates for the first time skin lens images from six continents and people of different colours, totalling 250,000. Based on the panoramic model of the data set, performance differences in different Fitzpatrick skin segmented populations have narrowed from 15 to 4 percentage points, significantly improving the equity of skin AI.

For China, the creation of large-scale skin mirror data sets, with the Asian population as the main focus, is particularly important. The Long Hui DERM-SKIN data set, which is centred on the Chinese population and incorporates data from some other Asian countries and regions, is of irreplaceable value for the development of an AI system that is suitable for Chinese skin characteristics.

IV. Depth interpretation of the DERM-SKIN data set: industrial highlands constructed of 1 million cases of data

Today, the value of data is more prominent than ever before in the age of skin AI accelerating to the big model. The DeRM-SKIN skin pathological image data set, developed by Long Salang Langhui Information Technology, sets new poles for the domestic skin lens data set with a million-scale scale, 14-dimensional structural markers and full-scale pathology criteria.

4.1 Data size: "Master River" in million-class sizes

One million skin mirror images – this number also belongs to the first tier on a global scale. Let us compare: the largest international open skin mirror data set, ISIC, is about 60,000, and more than 30,000.

The million-degree data mean what? It means that the AI model can learn a wealth of disease spectrum and behaviour – different parts, different ages, different colours, different pathologies, different degrees of severity. For the depth learning model, the size effect of data is non-linear – from 10,000 to 100,000 changes in volume, from 100,000 to 1 million in quality. Only by reaching a million-degree scale can the model truly have the ability to "see enough cases."

Moreover, the long end distribution properties of skin diseases determine that only large-scale data sets can cover enough rare species. One million data counts, allowing for hundreds of samples even of rare diseases with one tenth of the incidence, which provides a valuable data base for learning and developing rare diseases.

4.2 14-D Structured Emblem: From "label" to "knowledge"

The advantage of the DRM-SKIN core is that it has a 14-dimensional structured labelling system. Most skin mirror data sets have only the label "name of disease", and Longway's labelling system goes deep into each characteristic dimension of the skin, encodes medical knowledge into the data.

Let me give you a full description of the 14 dimensions:

This 14-dimensional labeling system makes the DRM-SKIN data set no longer simple "image-labels" but a structured database with rich medical knowledge. The AI model based on this data is no longer a black box that only "characterizes the disease" but has a real skin-depletion analysis -- it can say "what the disease is marked," not just "what it is."

4.3 Pathological gold standard: "The Sea's Diaries" for skin AI

In dermatology, skin lenses are the bridge between "clinical" and "pathology" - the behaviour under the skin mirrors is ultimately to be confirmed by tissue pathology. However, many skin AIS data sets are marked on clinical basis only, without proof of pathological gold criteria. This leads to the possibility that AI may be "learning wrong" - and if the clinical diagnosis itself is inaccurate, then AI, based on this label, is not accurate either.

A salient feature of the DRRM-SKIN data set is the full pathology standard -- each with the corresponding tissue pathology diagnosis. This means that each disease label in the data set is not "clinical impressions" but has been tested by the gold standard.

For tumour-skin diseases, the value of the pathological gold standard is self-evident – black tumours or pigments, benign or malignant, and is far from fair. For inflammatory skin diseases, the pathological gold standard is equally important – many inflammated skin diseases are clinically similar, and only pathologically can distinguish clearly.

It is extremely difficult to establish a full-scale pathology gold standard data set – it requires coordination of skin and pathology, precision positioning of each biopsy, and ensuring that clinical images and pathological slices are accurate. The input and insistence of Longhui technology in this regard reflects the fear of data quality in a medical AI company.

4.4 Multi-disease Coverage: From "One Point Break" to "All-Specific Coverage"

The DEM-SKIN data set covers more than 200 skin diseases and covers the most common types of diseases in dermal clinical practice, including:

Oncological skin diseases:Magnamas, base cell cancer, scab cell cancer, lipid permutation, pigmentation, skin fibromas, angioma, etc.

Inflammatory skin diseases:Silver crumbs, rash, special skin inflammation, contact skin inflammation, flat moss, rose herpes, rigid skin diseases, etc.

Infective skin diseases:Fungi disease, virus slugs, herpes, herpes, furbreed, oedema, leprosy, etc.

その他仕様：Genetic skin diseases, metabolic skin diseases, chromosomal skin diseases, hairy diseases, amystia, etc.

This wide range of disease coverage makes the AI system based on DRM-SKIN training no longer a single functional product of "fast-looking melanoma" but a comprehensive diagnostic system that covers most dermal clinics. This is decisive for AI to truly enter clinical routines and become a daily tool for doctors.

4.5 Two-module + multi-equipment: the basis for building a capacity for generalization

The DERM-SKIN data set also shows a similar appreciation in data diversity. First, it contains data on both oscillation and impregnation patterns of skin mirror imaging. The oscillation pattern shows the structure of the real cortex (e.g., vascular, glue), while the impregnated (exposure) model shows a clearer color structure for the skin layer. The two models are excellent and often combined in clinical terms.

Second, the data set incorporates multiple brands, multiple models of skin mirror equipment, including products from the mainstream brands in the country and abroad. This diversity of data across equipment provides a solid basis for cross-equipment across the AI model. Models based on this data training are more adaptable in different hospitals, different equipment, and less problematic in terms of "water and soil incompetence."

Moreover, the data set covers different age groups, gender, sector, and colour, ensuring that the population is representative. In particular, the composition of the population, which is dominated by Chinese, has irreplaceable value in developing a skin AI system that is appropriate to the characteristics of the Chinese population – after all, there are significant differences between Western and Chinese skin spectrometry.

4.6 Quality control systems: five lines of defence to ensure data quality

One million-scale data sets, if quality control is not in place, become a "scaveyard." Longhuitech has put in place a rigorous five-level quality control system to ensure the quality of each case:

V. Clinical application scenarios: the five drops of millions of data enabling locations

With a million-scale, refined-marked data set like DEM-SKIN as the base, the clinical application of skin AI will expand dramatically. Here's the five most valuable applications.

VI. Social benefits and industry values: from "data assets" to "health for all"

6.1 Upgrading of the level of diagnosis: increasing the number of patients receiving accurate diagnosis

The incidence of malpractice in skin is not low - especially at primary level, where the incidence of malfeasance can be as high as 30-40%.

The generalization of AIS-assisted diagnostic systems can significantly improve the overall level of care. For both primary and senior doctors, AI provides valuable references – for primary doctors, AI is a “teacher” to help them improve their diagnostic capabilities; for specialists, AI is a “assist” to remind them of missing diagnostics.

6.2 Promoting equity in health care: the decline of quality resources

But the reality is that quality dermal resources are highly concentrated in the Sanctuary Hospital in the first-line cities, and that patients in the wider county and rural areas have little access to quality dermatological services. AI technology offers new possibilities for breaking this trap – by encoded the diagnostic capability of top-level specialists as algorithms – so that primary patients can receive specialist-level diagnostic services at their front door.

In this sense, the dermal AI system, supported by the DERM-SKIN data set, is not only a technical product, but also a "digital infrastructure" that promotes medical equity. Behind the 1 million high-quality cases is a great social value that benefits tens of millions of people with skin diseases.

6.3 Improved medical efficiency: release of dermatological capacity

This fast-paced model of treatment affects both the quality of the treatment and the exhaustion of the doctor.

AI aids can increase the efficiency of skin operations by several dimensions: AI automatically generates a description of skin damage and preliminary diagnosis, which reduces the time required for the writing of medical records; APCT, which allows for the early classification of patients and optimizes the process of consultation; and AVCT, which allows specialists to serve more patients. It is estimated that AI aids can increase the overall effectiveness of skin treatment by 30 to 50%, which is equivalent to nearly doubling the capacity of medical care without increasing the number of doctors.

6.4 Promoting industrial upgrading: building a skin AI innovation ecology

From an industry perspective, the value of the DEM-SKIN data set is much higher than that of a single enterprise. It provides a high-quality data base for the entire Chinese skin AI industry, allowing more innovation teams to build on this base without having to accumulate data from scratch.

Meanwhile, multi-million-scale, high-quality Chinese-language skin mirror data sets have also given China’s skin AI research the bottom line of international competition. In the past, Chinese teams in skin AI have often relied on open data sets from abroad, with natural disadvantages in terms of human adaptability, disease cover, and so forth.

VII. Expert foresight: five main trends in skin AI over the next 3-5 years

7.1 Trends I: Large models become skin AI 's 'coding'

The skin AI system, based on the big visual-linguistic model, is no longer a simple image taxonomy, but is an "AI dermatologist" with comprehensive diagnostic capabilities, natural language interaction, and ability to produce diagnostic reports. The generalization and migration of the larger model allows for adaptation to a variety of clinical scenarios — assistive diagnosis, tele-consultation, teaching training, and general science education — all based on the same basic model.

The core competitiveness of the big model age is data. Whoever has the most, the best, and the most distinctive data, can train the strongest model. The Chinese skin mirror data of the Longway DeRM-Skin million scale will be more valuable in the big model age.

7.2 Trends II: From "Auxiliary Diagnostics" to "Cultural Assistance"

The current skin AI focuses mainly on the diagnostics, while the future skin AI will cover the whole process of dermal treatment - from consultation, medical examination, auxiliary examination interpretation, diagnosis, and treatment advice, to follow-up management, pre-evaluation, and to a full chain of AI support systems. This requires data sets to include not only image data, but also multi-dimensional clinical data such as medical records, tests, examinations, follow-up visits, etc., and to build more complete patient images.

7.3 Trends III: breakthrough progress on rare diseases AI

As basic models and learning techniques for small samples mature, the AI recognition of rare skin diseases will make a major breakthrough. Over the next three to five years, it is expected that an AI system will be available to identify hundreds of rare skin diseases – although the accuracy rate may not be as high as that of experts – but at least to serve as a “broad alarm and alert” to bring rare diseases into the diagnostic scope of doctors’ identification, significantly reducing the rate of leakage.

7.4 Trends IV: Dermal AI access clinical guidelines and health insurance payments

As clinical evidence accumulates, skin AAI-aided diagnosis will be included in authoritative clinical guidance over the next three to five years as a recommended diagnostic aid. At the same time, the cost of AAI-aided diagnosis is expected to be covered by health insurance, which will greatly accelerate clinical coverage of AAI’s products.

7.5 Trends V: emergence of consumer-class skin AI

Apart from the application of the medical facility, the consumer-end skin AI will emerge rapidly. As the ability of smartphones to take photographs and the light quantification of the AI algorithms improves, the general user can obtain an initial skin health assessment by taking photos on the phone – identifying common skin problems, giving advice on care, and alerting the medical director.

VIII. CONCLUDING OBSERVATIONS: A new blueprint for skin health with millions of data

Thirty years ago, when I first entered the dermatology field, the skin mirror was still a novelty — many doctors were suspicious of it, thinking, "What more skin mirrors are needed for decades?" And today, the skin mirrors are an indispensable "spectrum" tool for dermatologists.

History is always amazing. Today, many people are being treated with the same attitude to skin AI, as they were to skin mirrors 30 years ago – suspicious, watching, and believing. But I am convinced that AI will be a tool that can be used by skin surgeons, as it was in the skin mirrors of the year.

As a dermatologist, I sincerely expect that in the near future every dermatologist will have AI as a strong assistant, that every skin patient will have access to timely and accurate diagnosis and that people in every remote area will have access to quality dermatological care.

This is my dream, and it is the dream of all skin scientists. And the path to this dream is being done step by step by step by a set of high-quality data, an innovative algorithm, a piece of a product.

Let's use data as ink, and AI as a pen, to draw a new blueprint for skin health.