Skin disease diagnosis system using machine learning

Abstract

In Uganda today, skin disorders are widely spread amongst the citizens of whom some are unaware that they aren’t fully healthy. A skin disease can be defined as a specific or a particular type of condition caused by various factors that may include bacteria, fungi, hormonal imbalance amongst others and can lead up to several skin conditions such as dermatitis, hives, melanoma, ringworm or even worse skin conditions. There is a general problem of delayed diagnosis that can be attributed to the very few dermatologists available and the cost of such diagnosis. While many people’s first step involves going to a Google search bar, it can be difficult to describe what you are seeing on your skin through words alone. It is at this point that image processing using machine learning comes in to save the day. Our project, using the advancements in technology and Artificial Intelligence, has been able to quicken the diagnosis process and ease the dermatologists work.

Artificial intelligence (AI) has the potential to help clinicians care for patients and treat disease – from improving the screening process as is in breast cancer to helping detect diseases more efficiently as is in tuberculosis [5]. When we combined these advances in AI with other technologies, like smartphone cameras, we unlocked new ways for people to get better informed about their health. Methods such as feature extraction, segmentation, image enhancement etc. are the core of image processing. They efficiently recognize if the part of the body input into the system using an image is affected by the disease. They identify the color, texture and form of the diagnosed part. An inclusive study of three skin diseases i.e measles, melanoma and eczema and their detection are done in this project report.

After the predictive model was developed, it was tested against a few performance metrics i.e a confusion matrix, precision, recall and F1 Score and with the help of SciKit learns (a python library) these results were attained to later define the model as highly accurate. The model provides accuracy of 82.87% , precision of 84.87% and recall of 82.68%. The web app performance was also tested using load and stress tests using Jakarta JMeter and it was determined that it could handle at most 1000 users given the available system resources.