PASSION derm

Access to healthcare remains a pervasive challenge in numerous countries globally, primarily stemming from an inadequate number of specialized doctors.

The scarcity of dermatologists in many African and Asian nations notably obstructs local populations’ access to crucial dermatological care. With often fewer than one dermatologist per million patients in these regions[1], the situation reaches alarming proportions for the majority. The affected population, notably up to 87% among infants[2], inevitably extends its impact to adults. These consequences span physical, psychological, and socio-economic dimensions, particularly in cases of chronic illnesses.

The PASSION initiative aspires to deliver a realistic solution, alleviating this predicament and providing local populations with essential access to dermatological care.

[1] https://www.facebook.com/watch/?v=458123999804780,  [2] Kiprono SK, Muchunu JW, Masenga JE. Skin diseases in pediatric patients attending a tertiary dermatology hospital in Northern Tanzania: a cross-sectional study. BMC Dermatol. 2015 Sep 10;15:16. doi: 10.1186/s12895-015-0035-9. PMID: 26359248; PMCID: PMC4566193.

The PASSION project unfolds in three main phases.

Phase 1: Data Collection, AI Model Training

1.1 Data Collection of Globally Representative Skin Images

Existing open-source datasets of dermatologic diseases are unfortunately not representative of African or Asian populations, with skin pigmentation primarily falling within the “white” skin types I, II, III on the Fitzpatrick scale [1]. To address this, since 3 years we actively collect data (images and metadata of skin types IV-VI) to train/evaluate artificial intelligence models and to build a searchable atlas for medical personnel. Data collection is performed by dermatologists in various countries and locations, facilitated through a dedicated application. For each case, after obtaining consent from the patient for data collection for research purposes, images of the lesion and metadata are collected. By the end of 2023, >4000 images have been gathered in Madagascar, Guinea, Tanzania and Indonesia.

Quality assurance involves both manual (visual verification by board-certified dermatologists of the diagnosis for every submitted case) and automated checks using our SelfClean algorithm [2]. Data collection is performed strictly in accordance with local regulations.

1.2 AI Model Training

The project builds on our own developed AI infrastructure that we open source globally.

• A foundation model trained extensively in a self-supervised manner with over 2 million skin images serves as the basis [3] (Switzerland).
• A classification model based on the “vision transformers” architecture (ViT[4]), aims to identify four common skin disease families (atopic dermatitis, scabies, impetigo, fungal infection, others) (Switzerland). Two versions of the classification model are created:
  1. Model A is trained with >4’000 dark-skinned images that were collected so far.
  2. Model B is trained for comparison purposes with 16’000 white images.

Phase 2: Creation of the Teledermatology Platform and Infrastructure for Clinical Trials

2.1 Creation of the Teledermatology Platform

After a testing phase of several months in Madagascar and Tanzania, we were able to ascertain that teledermatology based on a webservice is insufficient in speed and ease of use. Thus, we use standardized forms within the ubiquitious application WhatsApp as the communication tool, leveraging its encryption and existing distribution in the population. We have achieved the integration via API of AI models, interaction through a chatbot.

2.3 Clinical Trials

We are already in planning of a clinical trials aim to validate the AI-powered teledermatology approaches. The trial will first take place in the hospital and in the same patients compare the accuracy of in-person consultations, teledermatology done by local dermatologists and AI-teledermatology. Subsequently, when an accuracy of  >90% is reached by teledermatology, the trial will be performed in a remote setting with true teledermatology.

2.1 Selection of Existing AI Models for Testing in Clinical Trials

The data collected for teledermatology will be used to evaluate our own developed classifier model (Model A, see above), but also allow to evaluate existing AI models in dermatology (Belle.ai, Google Lens, VisualDx, open-source models, etc.). Subsequently, we will be able to choose those with the greatest potential for the project in terms of accuracy, speed, and adaptability to different cases.

Phase 3: Finalization of the Platform and Deployment of the Best Solution, and Publication of Datasets and AI Models

3.1 Finalization of the Teledermatology Platform

Analyze the results of clinical trials and adjust the platform based on feedback and discoveries. Optimize the user interface and features to ensure an optimal user experience.

3.2 Open Source

Our aim is to open source all our efforts with the intent of reducing the load of dermatologic disease globally. We will publish 60% of collected cases, the remaining will be safeguarded as validation dataset for future competitions. We will also publish all AI models (general model and classifier), the code for the teledermatology platform, code for AI training, and their utilization.

Conclusion: The PASSION project aims to effectively integrate AI technology into the field of dermatology, using clinical trials to compare different approaches. The finalization of the platform and the deployment of the selected best solution depend on conclusive results from clinical trials, ensuring a practical and beneficial application of teledermatology and AI in the medical field.

The solution we propose is an AI-driven teledermatology platform revolutionizing dermatology care in regions facing a shortage of dermatologists. Skin diseases are usually addressed with conventional in-person consultations. These are associated with extended waiting times and travel costs, even if dermatologists are available. Teledermatology by human dermatologists can solve the problem of accessibility, but the number of local experts remains constant, making this approach unsustainable.

Indeed, our approach overcomes these significant problems even in remote locations with the sole condition of network coverage. At any given time, patients’ images can be uploaded and diagnosed by our AI-assisted teledermatology platform. Initially, a validation study will accompany these efforts to avoid subpar AI-results. In addition, medical personnel inexperienced in dermatology will be trained by managing the AI-guided cases. For ultimate safety, we will continuously provide a certified dermatologist overseeing the AI-driven teledermatology platform.