Responsible AI for your Skin
Skinopathy AI was developed and trained by Doctors and Privacy Advocates to protect you and your data
What is AI?
Artificial Intelligence (AI) is a tool, just like your cell phone is a tools
It is how an AI is built and how it is used that makes it morally right or wrong
What Data is Used to Train Skinopathy AI?
AI models are trained on data, and where AI developers find that data can sometimes be murky or unethical.
Skinopathy AI gets its datasets from four different places, and we make sure they are used ethically.
Patients give us explicit and informed consent to use their images/data.
The Canada National Research Council and Alberta Health Services have deemed our use of data to be ethical through their Research Ethics Board (REB) committees. Through that REB, we have been given access to other anonymized patient data.
Open Source repositories online that have been vetted by the community of data scientists and doctors and have been given special licenses to be used to test AI models.
We create “augmented datasets” that are artificially expanded through mathematics. These transformations help to create a more diverse training set for AI models, which can improve overall usage.
Is the AI Explainable?
There is a school of thought that says all AI should be explainable.
Think back to Grade 10 math when your teachers asked that you “show your work” to the right answer.
Same idea, but way more sophisticated and technical.
Skinopathy AI is explained in the following ways:
We published research papers, which you can see here, where we discuss the “nutrition level” of our training data and the limitations of our AI.
We provide clinicians and patients with the medical logic maps underlying the AI’s decision systems.
We provide novel saliency maps (a heat map that shows the regions most important for an AI), which explains the rationale on how and why an AI made an assessment.
Is there a Human-in-the-Loop?
There are many types of AI available right now. However, not all of them have a “Human-in-the-Loop.”
Human-in-the-Loop is a mechanism that makes an AI safer and more accurate by quite literally adding a human as part of the AI’s growth.
For example, having an expert confirm the AI results are actually correct and feeding that confirmation back into the training of the AI.
Skinopathy AI does this, but takes it even one step further by:
Asking trained doctors to visually verify that the assessment of the AI is accurate.
Using the results of blood tests and biopsies to confirm the accuracy of both the AI and the doctor and to also use those test results to train the AI.
Is There Bias in the AI?
Training an AI is just like training for anything else.
The student (the AI) is liable to be trained by a person who has a slanted view of the world, and that view can have a strong influence on the student.
And Skin Tone Bias is a very real concern in medicine. So, what is Skin Tone Bias?
It is the tendency to overlook specific qualities, consciously or unconsciously, when assessing skin conditions from People of Colour.
For example, it can be difficult for trained and experienced physicians to visually determine the amount of inflammation (whether through injury or chronic skin condition) on a patient with a darker skin tone due to colours blending together.
Skinopathy AI has shown:
It is able to detect the miniscule colour changes in a patient’s skin tone that a trained dermatologist cannot.
It has been shown to be comparable to other costly medical tools, like Laser Doppler Imaging devices.
Data
AI models are trained on data, and where AI developers find that data can sometimes be murky or unethical.
Skinopathy AI gets its datasets from four different places, and we make sure they are used ethically.
Patients give us explicit and informed consent to use their images/data.
The Canada National Research Council and Alberta Health Services have deemed our use of data to be ethical through their Research Ethics Board (REB) committees. Through that REB, we have been given access to other anonymized patient data.
Open Source repositories online that have been vetted by the community of data scientists and doctors and have been given special licenses to be used to test AI models.
We create “augmented datasets” that are artificially expanded through mathematics. These transformations help to create a more diverse training set for AI models, which can improve overall usage.
Explainable AI
There is a school of thought that says all AI should be explainable.
Think back to Grade 10 math when your teachers asked that you “show your work” to the right answer.
Same idea, but way more sophisticated and technical.
Skinopathy AI is explained in the following ways:
We published research papers, which you can see here, where we discuss the “nutrition level” of our training data and the limitations of our AI.
We provide clinicians and patients with the medical logic maps underlying the AI’s decision systems.
We provide novel saliency maps (a heat map that shows the regions most important for an AI), which explains the rationale on how and why an AI made an assessment.
Human-in-the-loop
There many types of AI available right now. However, not all of them have a “Human-in-the-Loop.”
Human-in-the-Loop is a mechanism that makes an AI safer and more accurate by quite literally adding a human as part of the AI’s growth.
For example, having an expert confirm the AI results are actually correct and feeding that confirmation back into the training of the AI.
Skinopathy AI does this, but takes it even one step further by:
Asking trained doctors to visually verify that the assessment of the AI is accurate.
Using the results of blood tests and biopsies to confirm the accuracy of both the AI and the doctor and to also use those test results to train the AI.
Bias
Training an AI is just like training for anything else.
The student (the AI) is liable to be trained by a person who has a slanted view of the world, and that view can have a strong influence on the student.
And Skin Tone Bias is a very real concern in medicine. So, what is Skin Tone Bias?
It is the tendency to overlook specific qualities, consciously or unconsciously, when assessing skin conditions from People of Colour.
For example, it can be difficult for trained and experienced physicians to visually determine the amount of inflammation (whether through injury or chronic skin condition) on a patient with a darker skin tone due to colours blending together.
Skinopathy AI has shown:
It is able to detect the miniscule colour changes in a patient’s skin tone that a trained dermatologist cannot.
It has been shown to be comparable to other costly medical tools, like Laser Doppler Imaging devices.
Our doctors believe the Hippocratic Oath extends to digital health
The Hippocratic Oath
The Hippocratic Oath’s most sacred tenet is to “First Do No Harm” and it is something we take very seriously.
That is why we have taken several measures to ensure patient data and privacy extends to their digital health profiles.
There are also regulations that govern how medical, and healthcare companies need to ensure patient safety in Canada. Below are a few that you should know about:
Our medical AI is a Class I device by Health Canada, which means it poses a lower risk to users (e.g. like a thermometer).
We have our Medical Device Establishment License (MDEL) with Health Canada, and that means our medical platforms are subject to audits and standards to ensure patient safety.
We are PHIPA and PIPEDA compliant and follow all healthcare regulations.
Skinopathy AI Models
EXAMPLE: Hyperpigmentation & Large Pores
Skinopathy AI dissects, pixel-by-pixel, the image you capture using your phone and provides objective data regarding your skin condition
This data can be used in many different ways, including product recommendations.
There are three core technologies that powers our AI for skin conditions
“Classifiers,” which is a neural network trained on millions of images, including images of specific skin conditions, such as skin cancers or hyperpigmentation.
We use trained neural networks to identify the condition of your skin, such as fine lines and wrinkles.
“Activation” Maps, which are neural network that reveals what an AI model “sees” within its neural layers.
These maps can tell you objectively where erythema (skin redness) is present due to acne, but similar maps can also tell us about the areas on your face that are sensitive or reactive to a particular product.
“Attention” Maps are information maps that integrates what an AI “sees” with what it pays attention to when performing its analysis.
These are the borders of a skin condition or analysis.
For example, this can be the features on your face that are not prominent to the naked eye but has been recognized by the AI as a feature that makes an individual look older than your actual biological skin age.
