AIforHer, the challenge combining artificial intelligence and humans to speed up the fight against breast cancer

When technological innovation meets therapeutic innovation, for the benefit of patients and healthcare professionals

Breast cancer treatment is improving all the time, but that’s not enough: we need to diagnose the disease more accurately and target treatments more effectively. With the arrival of new targeted therapies for patients with low or very low expression of the HER2 receptor, there is a clinical imperative to refine and make more reliable the diagnosis of this biomarker, which has become central to therapeutic decisions.

This demand for precision is matched by another upheaval: the rise of digital pathology and artificial intelligence algorithms capable of assisting pathologists in the detailed assessment of HER2. While these technologies are promising, their deployment in routine clinical practice still raises crucial questions: reliability, robustness, interpretability and acceptability to professionals.

It is at this juncture, between therapeutic progress and technological innovation, that a strong need is emerging on the ground: to ensure that the AI tools now available meet clinical requirements and are effectively integrated into medical practices.

To meet this challenge, a number of players are mobilising, drawing on their complementary strengths and their shared commitment to innovation for the benefit of patients:

GEFPICS, a network of excellence bringing together France’s leading experts in anatomopathology, oncology and biology, with an ambitious vision of AI as a lever for harmonising and improving diagnosis.
The Institut de Cancérologie de l’Ouest and its innovation accelerator, Impulse by ICO. The ICO is an expert cancer centre and a pioneer in digital pathology, with a routine infrastructure already in place, a dedicated AI unit, and recognised clinical expertise in senology (particularly in HER2-related issues). Launched in 2024, Impulse by ICO is the ICO’s oncology innovation accelerator. Its aim is to reveal and accelerate innovations that will improve the lives of patients and the day-to-day lives of those working to prevent, detect and combat cancer. The accelerator connects innovations to the real world, enabling ICO healthcare professionals to develop tomorrow’s oncology solutions, and innovative companies to de-risk their products by using real-life data and experiments to develop solutions in collaboration with the scientific and medical communities.
Two pharmaceutical companies at the forefront of the development of HER2-low and ultralow targeted therapies, committed to supporting innovative diagnostic approaches in line with new therapeutic directions.

Together, these partners are launching AI for Her, a call for solutions to identify, test and develop the best AI technologies applied to digital pathology in breast cancer.

The aim is clear : to enable every patient to benefit from the most appropriate therapy, as early as possible.

AI for Her

The pioneers of the greatest scientific discoveries and advances did not set out alone. Positive competition, team spirit and collective ambition in the service of a better tomorrow have been extraordinary drivers of development.
AI for Her aims to be an unprecedented challenge and a collective cooperative movement bringing together the worlds of technology and science to improve the care and treatment of breast cancer patients and help medical staff by bringing ever greater precision, objectivity and reproducibility to their diagnoses.
For a period of 6 to 9 months, anatomopathologists and start-ups developing an artificial intelligence solution capable of scoring HER2 (one of the Human Epidermal growth factor Receptor typologies playing a key role in breast cancer) at low and ultra-low expression levels, will be combining their expertise and focusing their energies on evaluating the performance and use of AI algorithms.

AI for Her is the very first challenge to enable AI algorithms to be evaluated in real life, at the heart of the hospital, thanks to the organisation of an entire cancer centre.

I want to apply

Scientific background to the challenge

Human epidermal growth factor receptors (HERs) play a key role in the development of many cancers. There are several types of HER receptor, we are focusing on HER2.

HER2 can be produced in abnormally high quantities by certain cancer cells. HER2 amplification or overexpression is detected in around 15% to 30% of breast cancers.

The level of expression of the HER2 receptor is now recognised as an essential biomarker because it has a dual value: prognostic, by providing indications of disease progression, and predictive, by making it possible to assess the probability of response to targeted anti-HER2 treatment.

Assessment of HER2 receptor expression is based primarily on immunohistochemistry (IHC), the reference method. This technique detects the presence of HER2 on the surface of tumour cells based on the completeness, intensity and proportion of cells expressing the marker.

The result is interpreted according to a four-level scale (from 0 to 3+), with clearly defined therapeutic implications:

Score 0 (HER2-0): Considered HER2-negative, with no indication for anti-HER2 targeted therapy.
Score 1+ (HER2-1+): Also classified as HER2-negative according to traditional criteria, but recently included in the HER2-low category.
Score 2+ (HER2-2+): An equivocal result requiring confirmation by in situ hybridisation (ISH) to detect possible genetic amplification of the HER2 gene.
Score 3+ (HER2-3+): Corresponds to strong overexpression of HER2 and constitutes a clear indication for targeted treatment with anti-HER2 therapies.

Long considered ineligible for targeted anti-HER2 therapies, tumours classified as HER2-0 and HER2-1+ are now the focus of growing clinical interest.

A breakthrough that has profoundly transformed the historical classification of breast cancer, which had previously been binary: HER2-positive (IHC 3+ or IHC 2+ with amplification) versus HER2-negative (IHC 0, 1+ or 2+ without amplification).

Since 2023, the recommendations of the American Society of Clinical Oncology / College of American Pathologists (ASCO/CAP) have included the HER2-low category to designate these intermediate tumours.

This change marks an important turning point, as it opens up access to targeted anti-HER2 therapies for patients previously excluded from these options.

More recently, a subgroup named ‘HER2-ultra low’ has emerged to describe IHC 0 tumours, however showing weak and incomplete membrane staining in ≤10% of cells.

Although this HER2-ultra-low category is not yet included in the official recommendations, the ASCO/CAP guidelines recognise the growing importance of distinguishing IHC 0 from IHC 1+, given the potential therapeutic implications.

Despite these promising therapeutic advances, these new classifications rely on IHC evaluation based on subtle morphological criteria, which can give rise to considerable inter–observer variability. This subjectivity can lead to misclassification of patients and compromise equitable access to recently available targeted therapies for tumours with low HER2 expression. The distinction within HER2-0 of the precise diagnosis between HER2-null and HER2-ultra low tumours therefore still remains a major challenge.

With this in mind, it is essential to have diagnostic tools capable of improving the accuracy and reproducibility of HER2 scoring for low and ultra-low expression.

Against this backdrop, companies specialising in artificial intelligence applied to anatomopathology have developed specific technological solutions (software/algorithms) designed to help healthcare professionals in this low and ultra-low expression HER2 scoring stage.

The democratisation of these solutions, which will improve patient care and help healthcare professionals in the future, will require a precise demonstration of their added value, in terms of pure HER2 scoring performance and their impact on the diagnostic performance of pathologists.

The Impulse by ICO call for solutions, AI for Her, has this objective and ambition.

Challenge methodology

For methodological reasons, the study was divided into two sub-studies.

Study A: assessing the performance of all the solutions

Comparative evaluation of the performance of several artificial intelligence tools for low and ultra-low expression HER2 scoring in breast cancer. No calibration is planned for this study of the entire cohort. At the end of this study, the two algorithms judged to perform best according to the defined evaluation criteria will be selected for Study B.

Study B: assessing the impact of the two selected solutions

To assess the impact of artificial intelligence on the diagnostic performance of pathologists in scoring low-expression HER2 in breast cancer. A calibration is planned for this study. 4 pathologists, divided into two groups (non-experienced and experienced) will individually evaluate all the slides from the ICO cohort without AI, and then with each AI selected in study A.

The study involved a retrospective global cohort of breast cancer cases consisting of :

200 samples from ICO patients. The 200 immunohistochemical (IHC) slides were prepared (staining and digitisation) by the ICO.
100 samples from patients outside the ICO. The 100 immunohistochemical (IHC) slides are prepared (staining and digitisation) under conditions other than those used at the ICO.

The performance of each solution (algorithm) will be assessed against two gold standards representative of current practices: (1) slide scoring by Labcorp*, and (2) slide scoring by experts from GEFPICS**.

Each algorithm will be compared with the two gold standards with the aim of comparing with current practices and not discriminating between the best gold standard.

The performance of the algorithms will be assessed mainly on the basis of the following quantitative and semi-qualitative criteria :

HER2 scoring, null, low or ultralow
Cell marking intensity
Completeness or incompleteness of labelling
Percentage of cells labelled

The performance of each solution will be assessed in accordance with the scoring grid below :

It should be noted that, in addition to the quantitative and semi-qualitative criteria, the project’s scientific committee will note, for each solution, qualitative information relating to the robustness, ergonomics and maturity of the solution, which will constitute, for each candidate start-up, concrete feedback from the experts. This qualitative information will nonetheless not be taken into account in the performance evaluation score.

*LABCORP: Laboratory Corporation of America Holdings is one of the world’s leading providers of medical laboratory analysis and diagnostic services. Present in more than 100 countries, the company offers a wide range of laboratory tests, from routine analyses to specialist examinations in genetics, oncology, infectious diseases and reproductive health.
Thanks to a strong capacity for innovation and the integration of the latest scientific advances, Labcorp supports healthcare professionals in the prevention, diagnosis and monitoring of pathologies, while contributing to clinical research. Its objective: to improve medical decisions and care pathways for a better state of health for patients throughout the world.

**GEFPICS: The Groupe d’Etude des Facteurs Pronostiques Immunohistochimiques dans le Cancer du Sein (Group for the Study of Immunohistochemical Prognostic Factors in Breast Cancer) is a French scientific group made up of experts in anatomopathology, oncology and medical biology, whose main objective is to standardise and improve the evaluation of immunohistochemical markers in breast cancer, and in particular: to establish technical and interpretative recommendations for the immunohistochemical analysis of biomarkers, to harmonise diagnostic practices, and to help improve the quality of diagnosis.

Procedures for candidates

(1) Having developed an algorithm for scoring low and ultra-low HER2 expression in breast cancer.

(2) This algorithm must be CE marked (or equivalent), in the process of being marked or at a level of maturity that allows it to apply for marking in the near future

(3) Having a functional algorithm with the IHC slides used in the study: staining with the Roche Ventana Automated Benchmark Ultra and Roche Ac 4B5; Scanning with the Leica Aperio GT450 and 3D HISTECH scanner.

(4) The results of the algorithm must be provided in a report in a usable format (csv, xls, etc.) enabling each slide and its associated performance criteria to be identified, as well as the average processing time per slide and the technical failure rate (number of slides on which the AI was unable to reach a decision in relation to the total number of slides assessed.

(5) Being able to provide, within the allotted time and with the results, the server logs to ensure that the results are autonomous.

Applications must be sent via the dedicated platform. The application must include all the documents requested:

A presentation document in PPT or Word format (max 10 slides) specifying the technical architecture enabling the algorithm to operate and in particular the type of algorithm used (which technology, proprietary or open source algorithm, etc.), how the model has been developed (in particular the quality and quantity of the training dataset), how the results are transcribed and visualised (in particular the intensity of labelling, whether labelling is complete or incomplete, the percentage of cells labelled and the HER2 score, etc. ; Screen prints and visuals of your viewer are recommended), a copy of the results report for study A, how you guarantee that the results provided by your algorithm are 100% autonomous, without human intervention or manual annotation prior to the results report, the time required to provide the scoring results for the entire cohort (study A) and any initial performance results obtained on data other than the training data.
The CE marking certificate (or equivalent) obtained or the certificate of deposit.

Projects must be submitted during the opening period of the call for projects and will then be collected and processed on the closing date of this period.

Applications will be analysed on the basis of the above-mentioned eligibility criteria and the completeness of the application. Applicants who meet these two conditions will continue the selection process and will then be able to demonstrate their solution at a Startups Demo Day in front of three GEFPICS experts, the industrial partner and all the ICO healthcare professionals involved in the project (pathologist, oncologist and AI expert). This Startups Demo Day is a unique opportunity to showcase your technology to those working in the field.

It has already been agreed that this Demo Day will enable GEFPCIS experts to evaluate the qualitative criteria of the algorithm that will be used for the final performance rating.

At the end of the demonstration, the members of the jury will decide which technological solutions will be selected to take part in the study on the basis of compliance with the eligibility criteria for the solution and assessment of the qualitative criteria..

Participating candidates will have unprecedented access to all the expertise and resources of the Institut Cancérologique de l’Ouest to test, debug and optimise their solution using real-life data. There is no financial compensation, but participants will have access to various prizes depending on their ranking in the AI for Her challenge :

NB. With the exception of the solutions ranked first and second, the results published will be anonymous. Each participant will be informed anonymously of his or her performance ranking in relation to the other participants.

***DMH: The Digital Medical Hub is a key player in digital innovation in healthcare, supporting projects from clinical trials through to market access, integrating the associated product and regulatory strategy. For AI for Her, the Digital Medical Hub is responsible for the methodological aspects of the study (drafting and registration of the protocol), as well as assisting the winner with an early meeting with the HAS.

Challenge schedule

*The launch of studies A and B is subject to contractual agreement with all stakeholders.

For Study A, the timeframe indicated is that of the complete study including statistical analysis of the results. The duration of the scoring of the algorithms is to be determined, but will be spread over a shorter period (a few days).

For study B, the timeframe indicated concerns the three scoring stages (without AI, with AI A, with AI B), each of which will be spread over a few days, with a month’s wash-out planned between each stage to ensure the objectivity of the results. The timeline could extend to the end of February, including the analysis of the final results.

You are a start-up developing an artificial intelligence solution capable of low and ultra-low expression HER2 scoring?

I want to apply

To go further…

Understanding the different types of breast cancer

There are several types of breast cancer:

Invasive carcinomas, which, depending on the anatomical origin of the tumour, include ductal carcinomas and lobular carcinomas.
In–situ (non-invasive) carcinomas, which also include ductal and lobular carcinomas.
Rare cancers: inflammatory breast cancer, Paget’s disease of the nipple, phyllodes tumours, breast sarcoma, male breast cancer, etc.

Each cancer is then characterised by a stage from I to IV, and by a grade from I to III. Stage 4 represents metastatic cancer.

Cancer subtypes can then be determined on the basis of the expression (or otherwise) of certain molecular biomarkers:

Triple-negative cancer: presents none of the three markers
Hormone-dependent cancer: the cancer cells have oestrogen and/or progesterone receptors. This is known as ER+, PR+ or, more generally, HR+ cancer; or (ER-, PR- or HR-) in the absence of these receptors.
HER2 positive cancer: in the case of overexpression of the HER2 protein. This is known as HER2+ cancer, which is the subject of our challenge.

**Additional sources and diagrams** : Understanding Your Breast Cancer Diagnosis: A Guide – OWise UK

Understanding the patient journey and the angle of the study

The screening phase is mainly carried out by mammography in primary care.

During the diagnostic phase, several consultations are carried out:

A consultation with a surgeon and a medical oncologist
A radiology consultation with one or more imaging procedures (ultrasound, mammography, tomosynthesis)
A tissue sample (biopsy), unless a previous sample has already been taken

The diagnosis is announced on the day or within a week if the biopsy had to be taken during the initial consultation.

Once the biopsy has been taken, the work of the anatomopathology laboratory begins:

Fix the tissue in formalin to preserve the cell structure and embed it in a paraffin block to stabilise the tissue for slide preparation.
Create a slide by cutting a thin slice of the paraffin block (a few microns), then staining this slice to highlight tissue structures* or specific biomarkers* (such as HER2)

For example, HES (Haematoxylin – Eosin – Saffron) staining is used for standard diagnosis, whereas IHC (Immunohistochemistry) staining is used to identify target proteins or antigens and therefore cancer subtypes.

In our challenge, we work on IHC slides.

The staining is then carried out using a combination of machine and antibody. The choice of antibody depends on the target.

In the case of HER2, we need to :

Use the Ventana HER2 4B5 antibody
Then scan the slide using a specific slide scanner (at the ICO: Aperio GT 450 – Digital Pathology). The choice of scanner and zoom (X20 or X40) will affect image quality. The digitised slide generates an image that can then be viewed by a pathologist on a computer screen. And it is from these images that the AIs will work to automatically determine the HER2 score.

**Source**: modelling of the ICO process

Q&A

No, the challenge only addresses AI technologies for characterising the HER2 status of breast cancer. Within this scope of technological solutions, only those that are CE marked or in the process of being marked are eligible to apply. This is due to the objectives of the challenge, which is to compare several solutions, so they must have a similar scope and technological maturity.

The challenge consists of 3 main stages: the Startups Demo Day, Study A and Study B. The Startups Demo Day will take place remotely (for full details, see another question in this FAQ). Study A only involves AIs and will therefore also take place remotely. Study B, on the other hand, will involve pathologists who will meet face-to-face at ICO sites to carry out the study.

In public communications, only the first two participants will be named. The other participants will be anonymous. The results will be provided to all participants, who will be able to know their position in relation to the others, anonymously.

The study is a retrospective study, so the patients whose data is used will not be concretely involved in this challenge. This study is being carried out in full compliance with the General Data Protection Regulation (RGPD). Under no circumstances will the data transmitted enable a patient to be identified. Each player who receives data will be contractually obliged to delete this data and to provide proof of this deletion.

According to ASCO CAP recommendations

The ICO’s investment in the winning startup is not part of the reward plan for this challenge. All the ICO’s equity investments via its investment fund are made within the framework of its Investment Program and follow a specific investment thesis. Each investment by the ICO is accompanied by a collaboration involving the ICO’s expertise and resources, with the aim of securing the first funds raised and accelerating the R&D roadmap with investors.

While the measurement of such an indicator is important and relevant, the scope of the study does not include measuring the carbon impact of the AI solutions selected. This choice is explained in particular by the fact that the scope of the study would not make it possible to deliver a value representative of the real carbon impact of the solutions. In order to avoid publishing inconsistent results, the choice was made to focus on the performance and usage criteria of the solutions.

These are two pharmaceutical companies with a close interest in HER2-low and ultralow targeted therapies. These two partners are providing financial support for this project and are fully committed to the reward scheme for the two best solutions, as well as to the scientific exploitation of the study results.

Contact

Morgane MENARD
Head of the IMPULSE by ICO Accelerator
morgane.menard@ico.unicancer.fr

Antoine GUYONVARCH
Innovation and digital health project manager
antoine.guyonvarch@ico.unicancer.fr