Page 162 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

generated images of lesions on darker skin tones are realistic and imitate how a true lesion on
that skin would look.

This approach could not only improve the accuracy of lesion diagnosis on underrepresented
skin tones but also serve as an efficient, proactive diagnostic tool to help prevent disease
spread. For example, during the Mpox outbreak, which was largely concentrated in the Congo
and surrounding regions of Africa, with only a few cases elsewhere (World Health Organization,
2022), the availability of images showing Mpox on darker skin tones could be used to generate
synthetic images for lighter skin tones. This could significantly enhance model accuracy and
offer a preventive diagnostic tool for identifying new cases of Mpox on lighter skin tones. This
example highlights that, for certain diseases, the need for data might even require generating
conversions from darker to lighter skin tones instead, as described above. Once the datasets are
produced/augmented, anyone can train different classification models on them. However, most
of the models are not optimized to run efficiently. EquiDermAI goes a step further to train the
model on the datasets produced (now with the skin tones equally represented) and quantizes
them in order to run efficiently with tinyML. This process ensures that the model can be run on
low-power devices with little storage and no internet connection, thus being more accessible
to doctors and other individuals in underdeveloped regions with resource constraints to ensure
proper diagnosis. Overall, EquiDermAI aims to address as many diseases with discriminatory
training biases as possible in order for everyone to receive accurate diagnoses. It also aims to
provide everyone, regardless of where they are in the world, with the proper efficient AI-driven
diagnostic solutions in order to ensure an equitable future in dermatology

Use Case Status: Ongoing—In active deployment and continuous development.

Partner(s): None

2�2 Benefits of use case

EquiDermAI improves the early and accurate diagnosis of diseases across diverse populations,
and this directly contributes to enhanced health outcomes and reduced mortalities, further
enabling such diagnostic tools even on low-power devices for those in resource-constrained
areas, reaching as many people as possible, supporting universal access to crucial healthcare
services.

This framework mitigates both racial and regional biases in the field of dermatological
diagnosis with AI, thus ensuring fair diagnostic performance across diverse skin tones and
underserved populations. This promotes equality in the delivery of such pivotal healthcare
services, addressing disparities in the access and quality of healthcare.

By leveraging tinyML and edge computing, EquiDermAI minimizes computational resources
and avoids reliance on cloud-based infrastructure, reducing energy consumption. It fosters
sustainable AI deployment, particularly in regions with limited digital infrastructure.

2�3 Future work

Currently, EquiDermAI is able to generate its own skin lesion training instances with
underrepresented skin tones that accurately mimic that of expert-validated images and
training models that can use these datasets and run on low-power devices to perform accurate
classifications. However, future works will focus on expanding EquiDermAI as an app for several

126