Bioderma Congress Reports EADO 2025

Chair: Josep Malvehy (Spain)

Speakers: Josep Malvehy (Spain) and Katrien Vossaert (Belgium)

I have to confess that I still use the good old method of manual photographic comparison in my practice. I’m a total has-been – it dates back to the 1970s. My image quality is very poor but I can probably detect changes of more than 2 mm.

To improve follow-up, I can now use digital dermoscopy (to obtain more sensitive and lesion-specific information), total-body photography (to detect new lesions in high-risk patients), or a combination of the two to reduce the number of benign lesions removed.

New technologies offer better image quality and shorter procedures.

Total-body photography makes sense in cases of atypical mole syndrome: in just one second, almost 100 images are acquired, providing information on the multiple lesions, whereas digital dermoscopy of each lesion is not feasible.

The sensitivity of the combination [total-body photography + digital dermoscopy] is very poor on initial acquisition (around 25%) but rises to 100% on subsequent visits.

Existing devices are not standardised. When they are subjected to the USAF microscopic optical resolution test, different results are obtained: it is therefore necessary to know the state of the various dermatoscopic structures useful for diagnosis in order to determine whether a particular machine will be able to distinguish between them.

2D and 3D photographic systems do not allow us to see vessels, the pigment network, lines, dots, small globules or bright white structures correctly. However, they enable us to see colours, asymmetries, changes in diameter and pigmentation, pigmentation contrasts and pseudo-network aspects on the face, in order to target lesions for subsequent observation by dermatoscopy.

The latest full-body scanners use liquid lenses that enable the focus to be changed in just 1 ms, thereby greatly increasing the resolution.

There are also automated systems that combine total-body photography and dermatoscopic image-taking, which in around 10 minutes allow us to observe around 40 lesions dermoscopically with the same resolution as with a manual dermoscope.

Artificial intelligence (AI) tools can currently be used to segment lesions or dermatoscopic images, classify the morphology of individual lesions, detect changes, calculate melanoma risk scores and identify areas of interest or lesion similarity using explainable AI.

Explainable AI theoretically allows the user to check how an algorithm reached an answer. However, no authority is currently in a position to validate explainable AI tools!
As far as medical device legislation is concerned, obtaining the CE standard is legally necessary but not scientifically sufficient. Healthcare AI-based devices with machine learning must fall in Class II, a, b or III.

These tools have not yet been properly validated in terms of their clinical performance. They enable the patient’s phenotype to be determined and the number of lesions to be counted in a reproducible way.

The models are able to perceive changes in symmetrical or asymmetrical growth, focal or overall hyperpigmentation, colour change or the appearance of new lesions.

However, it is important to be aware of their limitations despite their good sensitivity and specificity. The few studies comparing devices, software and algorithms show that there are many false positives and false negatives, and that for the same patient the answers will be different depending on the device used. Unfortunately, the detailed performance of each algorithm is not available, which means that performance cannot be compared for different algorithms.

To test the performance of algorithms, Joseph Malvehy’s team artificially modified naevi. Sensitivity was good – close to 96% for 3D devices – but there were many false positives: positional, specific location, clothing, hair. However, it is interesting to note that melanomas had a significantly greater amount of change than benign lesions, and that the decisions made by the dermatologist assisted by the tool were more relevant.
In terms of acceptance by the public and users: patients generally have confidence in these new technologies, provided they are supervised by the dermatologist, and more novice practitioners are more likely to trust the results provided.

In the specific population of patients with atypical naevi, the use of this technology is particularly attractive because it reduces the number of lesions to be treated (NNT) to remove melanoma to less than 3:1.

These images are interpreted based on the clinical and phenotypic information available and the genetic or mutational profile where appropriate.

In the absence of a medical context explained by the dermatologist, almost 50% of melanomas will not be identified by the system.

The integration of these multimodal data by AI solutions requires prior training with good-quality datasets annotated with the necessary information.

Doctor K. Vossaert, from Ghent, Belgium, told us about her very interesting experience using 3D total-body photography in the general population and for monitoring high-risk patients.
Observing that the current 2D devices available on the market are ill suited to private practice because they are too costly in terms of staff time and space, she presented data from one year’s use of a 3D device by two dermatologists and four interns.

Because of the high cost, which would be insurmountable for an isolated practice, the device was installed in a stand-alone unit, with direct access for the patient or on referral, and with support provided only by an assistant who collected the necessary information. It performed dermatoscopic acquisition of lesions signalled after total-body photography by the algorithm via the pie chart, examining the skin folds, palms of the hands, soles of the feet and scalp. Each acquisition was reviewed by the dermatologist, who produced a report and recommendations. Any lesion analysed as suspicious at the end of the examination constituted a reason for referral for a face-to-face consultation. AI assistance was provided at two levels: selection of lesions after total-body photography and assistance in examining suspicious lesions using the malignancy score.

The very high sensitivity of these exams must be balanced by the dermatologist’s analysis, particularly for solar lentigines, blood vessels or tattoos, for example. The device therefore serves both as a diagnostic aid for high-risk patients and as a sorting aid for the general population.

Over one year of screening scans, 3,228 patients were photographed, including 64% with direct access and 36% who had been referred. This was not, of course, an organised, systematic screening programme in the strict sense of the term. Among the 2,059 patients who consulted spontaneously, high-risk patients were identified when there was a personal history of melanoma, or more than 100 naevi, including one atypical naevus. They represented 15% of patients.

Low-risk patients (59%) had to have fewer than 50 naevi and no personal history. The remaining 26% represented the median risk. In the low-risk group, 10% had suspicious lesions and the malignant lesion detection rate was 5.2% (i.e. 63 lesions for 1,208 patients within one year). The medium-risk group also had 10% suspicious lesions and a malignant lesion detection rate of 4%, i.e. 22 lesions for 545 patients. In the high-risk group, 18% of lesions were suspicious, with a detection rate of 4.2%, i.e. 13 malignant lesions. The total number of melanomas detected was one in the high-risk group, six in the medium-risk group and seven in the low-risk group.

The average detection rate of 4.8% is very interesting compared with data published 10 years ago in the literature, ranging from 0.8% to 3.2%.

The number of lesions that need to be removed to diagnose a malignant lesion is generally two, and three in the case of melanocytic lesions.

Seven melanomas were invasive, and the other seven in situ. Of these 14 melanomas, nine were incidental findings for the patient. The six other patients with melanoma had consulted a screening centre due to the lack of an accessible dermatologist. 1,590 of the 3,228 patients scanned did not require a subsequent dermatological consultation (all high-risk patients were referred, as were patients with suspicious lesions). In particular, 89% of the patients in the low- and medium-risk population did not require a consultation.

A comparative study between the automated total-body examination and the unassisted examination has also been planned.

Most melanomas occurred in the low-risk group of patients, due to their sizeable mass. For this reason, self-inspection is still recommended in this group, to look for new lesions or changes in existing lesions, with the CT scan potentially serving as a reference base.

The cost of the examination (around €100) was borne by the patient. A cost-effectiveness study is also required.

This model of fixed, expensive devices in practices may seem fragile in the face of proposals for mobile apps made to available to patients on their phones. However, while home-based solutions can reach certain groups who are unwilling or unable to consult a doctor, they are still limited by the quality of the images, the difficulty of autonomous acquisition, and the cost of the analysis required after the images have been taken.

Chairs: Julia Winkler (Germany), Allan C. Halpern (United States), Konstantinos Liopyris (Greece)

Speakers: Harald Kittler (Austria), Allan C. Halpern (United States), Josep Malvehy (Spain), Peter Soyer (Australia), Konstantinos Liopyris (Greece), Julia Winkler (Germany)

Introduction to current and evolving trends in AI

Any symposium dealing with artificial intelligence (AI) is always a source of wonder and panic for me. Changes are always asymptotic. As I write these lines, the trend is towards transformative neural networks or transformer models. These are types of neural network architectures with no sequential structure and which use semi-supervised or unsupervised learning. They are the backbone of foundation models such as Le Chat and ChatGPT.

Foundation models are machine learning models trained with generalised, unlabelled datasets and capable of performing various tasks such as natural language processing (NLP), answering questions and classifying images.

A foundation model has now been created specifically for dermatological use, to interpret a variety of image sources. The aim is to use it for screening, diagnosis, data analysis and prognosis purposes. For example, the user can identify changes over time in dermatoscopic images. From a dermatoscopic image, it can also generate a prediction about the characteristics of a given melanoma, which will be used to select an adjuvant or neoadjuvant treatment.

These dermatological foundation models are trained with datasets (sets of thousands or even millions of labelled images of dermatological diseases). As an example, since 2016 the International Skin Imaging Collaboration (ISIC) has been organising challenges aimed at improving the accuracy of melanoma diagnosis compared with other skin lesions. In 2024, hundreds of thousands of images were analysed and made available for more than seven different diagnoses. Databases are currently being enriched with images from a variety of sources (clinical photographs, dermatoscopy, histopathology, molecular biology, etc.). The current points of attention needed to minimise bias in dermatological image analysis are lesions that are underrepresented in databases, variations in skin colour, hair, and also markers placed on photos.

In Europe, explainable AI is required by legislation for AI in healthcare. This refers to artificial solutions for which the deep learning model can be understood by humans. While it may seem appealing, because it is more reassuring, it is currently mainly used to explain AI errors, but it does not provide any greater accuracy.

While it is now accepted that diagnoses are established more accurately by AI algorithms than by clinicians, the unresolved question is what to do with these new answers. People remain the holders of the information, and they choose when, where, how and to whom it should be transmitted. They must choose the useful fields of application for AI.
As far as training is concerned, various studies have shown that the information provided by AI is of benefit to less experienced users.
In research, AI is already being used in prospective real-life studies, such as the one where, in 2023 in The Lancet Digital Health, Menzies’ team showed that an algorithm was as good as a specialist at diagnosing a suspicious pigmented lesion using remote expertise on a mobile phone.

Doctors are also responsible for optimising the responses provided by AI. In dermato-oncology, for example, a false negative in the diagnosis of melanoma has far more serious consequences than a false positive. Positioning this cursor is just as important as increasing accuracy. It is therefore crucial to use feedback loops in therapeutic decision support models, using asymmetric penalties, for example.

AI tools can also select relevant information flows, for example when sorting requests for remote expertise.

Current state of imaging in dermato-oncology

Digital medicine, which includes the use of AI, is a $57 billion market in 2025. It includes, in no particular order: remote expertise and consultations, diagnostic applications for patients and healthcare professionals alike, and various advances in imaging. Although dermoscopy is now a routine part of dermatology practice, other imaging techniques are still less readily available, such as optical super-high magnification dermoscopy (OSHMD), which enables 400x magnification of structures similar to those seen under a microscope.

Total-body 3D imaging enables individual lesions to be identified by their unique coordinates in three dimensions. Once again, the way in which this technology is used has yet to be perfected, since the latest publication by the team of H.P. Soyer in JAMA Dermatology has just shown that for practitioners with little experience and without the benefit of AI, melanoma screening was less effective using the 3D system (35%) than in the control group (64%).

Deep skin imaging (confocal microscopy and LC-OCT), which aims to increase diagnostic accuracy beyond dermatoscopic images, currently seems to be more widely used in Europe than in the United States.

Current state of AI marketing and implementation in dermato-oncology

The development of AI tools for healthcare is complex and currently goes through a number of stages: first of all, the problem to be solved must be defined and the programme designed; then comes the data collection phase to develop the model; the pre-clinical phase for technical validation; and the real-life clinical validation phase. The next stage is to obtain CE marking and, more recently, achieve compliance with the European AI Act (applicable from August 2025). The product must then prove its qualities in terms of cyber security, integration into the IT system and ethics. Using explainable AI is also a requirement. Then comes the phase of commercial use of the system, with implementation in daily practice, evaluation of the product and issues of reimbursement by health systems, as well as ongoing evaluation and post-market monitoring.

In 2022 in JAMA Dermatol, the ISIC for AI published the list of items needed to evaluate imaging devices using AI in dermatology.

The key point is the dataset used to train the algorithm. The quality of the dataset is paramount, and manufacturers have a duty to be transparent about its origin. The algorithm must also be assessable and accessible. The product must also have a defined target audience (healthcare professionals, patients, etc.), an end user and a defined use.

There are several devices available with Class IIA CE marking, but in light of recent studies (Rotemberg, JAMA Dermatology, 2024, etc.), the smartphone apps currently available for melanoma diagnosis have not shown any clear value in real life, due to the lack of sufficient robust and accessible data.

Class IIa CE marking is not a guarantee of sufficient scientific quality! And, as in the pharmaceutical industry, proof must be provided through clinical trials.

Artificial intelligence remains a tool: the assistance it offers enhances the user’s intelligence. Studies have also shown that patients and doctors are more likely to trust a dermatologist aided by a machine than a human or AI alone.

Medical-legal and privacy issues related to the use of AI

The speaker discussed the ACEMID project, a network of three Australian states using fifteen 3D total-body photography devices, half of which are located in rural areas. The study has reached 16,000 visits from 7,500 participants; AI tools will soon be incorporated. The data currently being collected include 3D imaging and dermoscopic images, clinical data and patient questionnaires, Melanoma Institute Australia’s melanoma risk, genetic and proteomic data, data from Australian cancer registries, virtual slides and histological reports.

The use of these computerised imaging techniques in dermatology raises questions in terms of legislation and data confidentiality. Discussions are currently underway to issue recommendations for data governance (classification, security, anonymisation and retention of data).

Limits to the use of AI in dermato-oncology

The applications of 3D total-body imaging are not limited to melanoma. It can be used to monitor all kinds of dermatological diseases (assessing responses to treatment, calculating PASI and BSA scores, etc.). Errors in the interpretation of images by the AI algorithm can arise from changes in the patient’s morphology or position, areas of hair or tattoos or interpositions, changes in position, a segmentation problem or simply a lack of reproducibility.

Opportunities and challenges of AI in dermato-oncology

It is an established fact that convolutional neural networks outperform humans when analysing images of pigmented skin lesions, even more so when the user is inexperienced. Currently, for a wider range of lesions, neural networks are as good as people, even when it comes to doubts about lesions that are difficult to diagnose, such as facial lesions!

However, AI performs poorly for mucous and nail lesions, as well as for vascular and adnexal lesions, lymphomas and all rare tumours.

Chairs: Christian Blank (Netherlands), Georgina Long (Australia)

Speakers: Georgina Long (Australia), Christian Blank (Netherlands), Alexander Eggermont (Netherlands), Julie Stein-Deutsch (United States)

Clinical development of neoadjuvant immunotherapy in melanoma

Perfect is the enemy of the good in the realm of neoadjuvant treatment. A different immune response is obtained when immunotherapy is administered neoadjuvantly or adjuvantly.

Neoadjuvant treatment for melanoma is currently administered six to eight weeks before surgery, and the response is then assessed on histology. A complete histological response is defined as a maximum of 10% residual tumour. A partial response is defined as 10 to 50% residual tumour and a non-response as +50%. The sum of the complete and partial histological responses constitutes the major histological response.

The key SWOG and NADINA studies demonstrated the efficacy of neoadjuvant immunotherapy and enabled major responders to be exempted from further treatment. Event-free survival curves overlapped in the two trials, with SWOG (HR 0.58) showing a 42% reduction in the risk of events (progression, recurrence or death from any cause) and with NADINA (HR 0.32) showing a 68% reduction. These results are gradually leading to treatment and reimbursement authorisations for these drugs worldwide.

A pooled analysis of neoadjuvant checkpoint inhibitors showed that combinations (with anti-CTLA4 or anti-LAG treatments) were superior to anti-PD1 treatment alone in terms of event-free survival. The pooled analysis also showed that 61% of major histological responses were achieved in combination, compared with 50% with neoadjuvant anti-PD1 treatment alone. These patients, who therefore did not undergo any further treatment, had a lower risk of toxicity, better quality of life and less of an economic impact. For BRAF-mutated patients who do not respond, this non-response can help guide the choice of adjuvant treatment with targeted therapy.
A major histological response augurs well, whatever the adjuvant treatment adopted. The overall survival of patients who received a neoadjuvant checkpoint inhibitor was better than that of patients who received adjuvant anti-PD1 therapy in the absence of a histological response.

Lastly, monitoring can be extended to six-monthly for patients with a major response. Histologically, the pathological response is estimated based on the percentage of viable tumour and the percentage of fibro-inflammatory territories (and secondarily of necrosis and melanophages). As far as pathologists are concerned, macroscopy of lymph nodes after neoadjuvant treatment is reduced by at least half under the new recommendations.

In terms of surgery, the results of the PRADO study (where, after removal of the index lymph node to establish the degree of histological response, no additional dissection is necessary in the event of a major response) will need to be confirmed by the ongoing MSLT3 study.

The pooled SWOG + NADINA results show that patients with high interferon gamma scores have a better fate. These good responder profiles could therefore be treated with neoadjuvant anti-PD1 therapy alone.
The NéoIreni trial is about to begin, with the aim of determining whether neoadjuvant escalation (PD1-ipilimumab-relatlimab) can be reserved for patients considered to be future non-responders to anti-PD1 monotherapy, using a Neopredict composite baseline test.

Next steps towards fully personalised neoadjuvant treatment for melanoma

There are several extremely interesting avenues to explore with regard to biomarkers:

The five-year results of the PRADO study show that the 60% of patients with a major histological response have a specific survival rate for melanoma of 98%, so any aggressive strategy should be discouraged.

On the other hand, for patients with a low interferon signature, treatment escalation with a high dose of ipilimumab/nivolumab is recommended, with the aim of achieving a compensated histological response and an improvement in relapse-free survival curves.

In patients who remain with a low interferon signature after neoadjuvant anti-PD1 treatment, the addition of ipilimumab increases the signature and, at the same time, the anti-tumour response. In the group of patients with a low interferon signature, the response rate can even be improved by adapting the treatment, without increasing toxicity.

Another patient profile that is a candidate for therapeutic intensification is the group whose tumours express PD-L1 weakly and soluble LRG1 strongly.

LRG1 is a glycoprotein that increases neovascularisation by acting on the TGF-beta signal. It is induced by IL-6 and hepatocytes and plays a role in the diffusion of micrometastases. Expression of this molecule is a poor prognostic marker for response to immunotherapy, in the same way as the interferon signature, but also for the overall outcome of the patient (as clearly demonstrated in the PRADO and OpACIN-neo patient cohorts).

Stress-related adrenergic signals reduce the immune response. Patients exposed to stress have poorer response and event-free survival rates, and poorer metastasis-free survival.

Activated T cells express more surface receptors for adrenaline, and their stimulation exacerbates immune exhaustion. Blocking beta-1 and beta-2 adrenaline receptors with a beta-blocker such as propranolol improves patient outcomes.

On the other hand, it appears that patients subjected to stress have a lower incidence of Grade III and IV immuno-induced toxicity.

What can we learn from other cancers?

“More cures, shorter treatments, less surgery” seems to be the new paradigm when it comes to neoadjuvant treatment for all types of cancer. This talk set out a number of examples to be followed for other diseases.

Cutaneous squamous cell carcinoma is a tumour with a very high mutational load. The New England Journal article by Gross’s team in 2022 showed a complete + near-complete response rate of 63%, which was underestimated by the imaging evaluation.

The arrival of subcutaneous anti-PD1 formulations (nivolumab-pembrolizumab-cemiplimab) will make their neoadjuvant use even easier.

For head and neck cancer, there is also a correlation between the pathological response (complete or near-complete) to anti-PD1 treatments, alone or in combination with ipilimumab or relatlimab, and cumulative survival (Li, Cancer Cell 2025). We are awaiting communications concerning Phase III studies on combined anti-PD1 and anti-LAG-3 treatment.

Neoadjuvant anti-PD1 treatments (pembrolizumab) are also proving useful in combination with conventional cytotoxic chemotherapy in the treatment of triple-negative breast cancer.

There is a significantly higher rate of complete histological response. Publications from 2024 showed an impact on overall survival following the authorisation of pembrolizumab as a neoadjuvant treatment. Use of the neoadjuvant nivolumab/relatlimab combination without cytotoxic therapies also shows an increase in major histological response. 
For lung cancer, the combination of nivolumab plus chemotherapy as neoadjuvant treatment for locally advanced resectable tumours shows a 10-fold greater complete histological response than chemotherapy alone, even though trials have only shown a significant improvement in overall survival for tumours expressing PDL1.

Neoadjuvant pembrolizumab/chemotherapy outperforms adjuvant pembrolizumab, which in turn outperforms chemotherapy alone, this time with a gain in overall survival!

Neoadjuvant immunotherapy is, of course, effective at treating cancers caused by microsatellite instability (MSI). The NICHE 1 study showed 100% histological response, including 60% complete response in dMMR colorectal cancers.

Studies testing the anti-CTLA4 + anti-PD1 combination (botensilimab + balstilimab) show response rates of around 30% in cancers with repair anomalies and 93% in MSI+ cancers, all without recurrence at one year. For bladder cancer, which also has a high mutational load, neoadjuvant durvalumab plus chemotherapy shows a benefit in terms of event-free survival and overall survival and drastically reduces the number of cystectomies required. Reducing morbidity will be the main objective of pivotal trials.
Neoadjuvant anti-PD1 therapies have also been shown to improve survival in recurrent glioblastoma, even after a single dose.

Making pathological response criteria more effective and pan-cancerous

Since 2020, there has been a histological anti-PD1 response score for all tumours combined. The speaker described the three components of the regression bed: residual tumour, necrosis and regression.

The histological response after adjuvant treatment is the equivalent of RECIST by imaging in patients with metastatic cancer.

The establishment of these scores requires that pathologists draw up new recommendations for macroscopic management and report writing, to be able to reduce this new workload in the future.

The latest harmonisation presented at ASCO in 2024 shows that the results are reproducible regardless of the type of cancer and the type of sample (primary tumour, sentinel lymph node, biopsy), both for the estimated percentage of residual tumour and for the percentage of regression.

The results of the lymph node or tumour analysis following neoadjuvant treatment are therefore given for three parameters (residual tumour, necrosis, regression), expressed in 10% increments, with mention of complete histological response, major histological response or near-complete response, but there are no current recommendations for partial responses or lack of response at this stage.