Inspiration

In the news, there has been a lot of talk about the dermatological symptoms associated with COVID-19 infection, even if the individual is otherwise asymptomatic. These symptoms may include rash, blisters, or itchy hives. While researching this topic, our team began to realize that skin conditions related to COVID-19 may look different between patients with fair skin and those with darker skin. A paper by Lester et al demonstrated that in a systematic review of pictures in scientific articles describing skin manifestations associated with COVID-19, 93% (120 out of 130) were taken with patients with the three fairest skin tones (Types I-III). 6% (7 out of 130) showed patients with Type IV skin, and there was no representation of the darkest skin tones (Type V and VI). This can lead to cognitive biases that contribute to underdiagnosis of COVID-19 infection in patients with darker skin that are otherwise asymptomatic.

This issue isn't just limited to COVID-19 though. Turns out, many other skin conditions that present differently in patients with dark skin compared to those with fair skin.

The following is an example of atopic dermatitis in an infant with dark skin compared to one with fair skin.

Image of differences between fair and dark skin in atopic dermatitis

If the right image is what medical students, nurses, and physicians see in textbooks and medical journals, you can imagine how easy it would be to misdiagnose the patient on the left!

Furthermore, 47% of dermatologists report insufficient exposure to patients with darker skin during their training, which directly impacts the quality of patient care and contributes to poorer health outcomes in minorities. Our team aims to resolve this issue by creating a web app that increases the representation of dark skin in medical databases, journals, and textbooks.

What it does

Our team's web app aids physicians and other healthcare providers in uploading pictures of their patients’ skin lesions, creating an open-source database of skin conditions in patients with darker skin tones. We then plan to use this database in downstream machine learning algorithms, including training a DCGAN to generate even more examples of skin conditions in darker skin tones. Our goal is to increase the visibility of skin conditions in all skin tones and remove cognitive biases that contribute to poorer health outcomes in minorities.

How we built it

  • low fidelity wireframe using balsamiq

  • high fidelity wireframe using Figma

  • Firebase Authentication

  • Image generator using tensorflow DCGAN

Challenges we ran into

Because there are no established datasets containing images of skin conditions in minorities or darker-skinned individuals, we had trouble figuring out a way to train the machine learning models.

We are also having a lot of issues with DCGAN.

Accomplishments that we're proud of

We have completed the UI design and built a working web app with login, sign up, image uploading and linked to the neutral style transfer, which allows conversion of Caucasian skin patches to darker skin.

Also, we have made a separate skin classifier that would aid skin condition diagnosis for doctors.

What we learned

We learnt a lot on the current problems and challenges faced by people of colours, different skin conditions, and DCGAN for generation of realistic fake images.

What's next for PixGen

We aim to eventually have our web app address more diseases in which the appearance of skin looks different between fair-skinned and dark-skinned individuals. These diseases may include, but are not limited to:

  • Lyme Disease

  • Eczema and Psoriasis

  • Kawasaki disease

  • Acne

  • Spider bites

  • Cancer

Built With

Share this project:

Updates