# Processing with AI ## Exploration : IA & Ethics Name: >**Chloé Tabone** > Subject: >**Detect dermatological problems using Computer Vision** >[TOC] ## Design brief ### Biais If we don't source our dataset with enough rigor, the following biais might appear: >1. The first step in detecting a skin disease is to be able to detect something not normal on the skin. Therefore our model has to be able to detect skin issues on different types and colours of skin, not only clear ones. >2. The biais is not only about colour but also about age when it comes to skin issues. Indeed with age the skin can become less smooth which makes tiny things less easy to spot especially for a machine. >3. Finally there can be an unintentional bias coming from the use of cosmetics or not. A computer won't be able to know if someone has cosmetics on it's skin or if it is something natural. When a dermatologist will think abouot asking this question most of the time. > We will ensure that our model is not biaised by: >1. Sourcing our data from different countries and population. The goal of this ork is to get a sample as diverse as possible to train it with the most different skin colors and textures. >2. Making sure our data take into account the age bias will be the biggest challenge. It's not only about computer vision but about how the pictures are taken. >3. The cosmetic aspect of bias wil be the easiest to eliminate as you can simply display a caution message saying that the program will only be effective if you are not wearing any make up, foundation, etc ### Overfitting >It's important to understand that skin diseases are not the same according to skin types, ages, and countries. They involve thousands of parameters. >It is therefore very important to focus not only on the diseases treated on your population but to train it to recognize as many different things as possible > We will make sure our model does not overfit by > Checking the accuracy of our model on different poopulations. We have to be careful not to train our model to detect only diseases such as melanomas for light skin but also include darker skin issues such as depigmentation for example. ### Misuse >We have to remind ourselves that our application could be misused by people who think they have an issue to avoid completly going to a dermatologist. This is an issue because the program will never be able to ask background questions etc as would a true dermatologist. >This program can be use as an help or a first step for detection or even when something is suspected to do a follow up check up without having to go to a doctors. >This program shouldn't be usable by everyone without any restriction. ### Data leakage >In a catastrophic scenario, where all of our training dataset were stolen or recovered from our model, the risk would be the conservation of the medical secrecy and confidentiality. >Medical informations are particularly sensitive so we have to make sure that they are correctly used and stored. ### Hacking > If someone found a way to "cheat" our model and make it make any prediction that it want instead of the real one, the risk would be that people could be treated for diseases they don't have with major health repercussions.