Diagnosing Skin Cancer Using Artificial Intelligence and Machine Learning

Artificial Intelligence (AI) and Machine Learning (ML) have many applications in the healthcare field. A lifesaving way in which these futuristic tools can be used is to diagnose skin cancer. I developed AI & ML models that can diagnose 7 different forms of skin cancer just from a skin lesion image. My goal was to enable people to upload an image of their skin lesion, and the model will generate a diagnosis for them and a percentage for the diagnosis’ accuracy. Thus, anyone can quickly and easily receive a precise diagnosis. I first researched about skin cancer and found that 1 in 3 cancers are skin cancers, indicating its prevalence. In certain communities, access to professional healthcare is scarce, depriving patients of quality care. To alleviate this, I created a website using AI & ML where people upload a picture of a skin lesion and obtain a diagnosis. Since people upload their own pictures, I had to consider that these images are likely not professional images, but the model should still diagnose it. Hence, I performed data augmentation on my dataset of skin lesion images. I made duplicates of my dataset and manipulated them by flipping, blurring, resizing, or zooming them using OpenCV. I then created numerous machine learning models such as K Nearest Neighbors (KNN), Convolutional Neural Network (CNN), Grid Search CNN, and Transfer Learning, to determine which model best diagnoses the different skin cancers. I evaluated the models using the Receiver Operator Curve (ROC), which shows the relationship between a model’s true positive and true negative rate. To interpret the curve, I used the Area Under the Curve (AUC) metric, which compares the model to one which randomly guesses. Additionally, I plotted Confusion Matrices to view a detailed configuration of each model’s performance. After evaluating, I found that my Transfer Model performed the best. For my Transfer Learning model, I used Keras’ VGG16 Machine Learning model as the base, and added my own layers of neurons to it. I further improved this model by considering different skin tones, since a bias in the dataset and model can be dangerous to those who use it. Thus, I trained my model on images of multiple skin tones. I then deployed this model to a website I created using JavaScript and HTML. My design allows people to visit my website, upload an image of their skin lesion, and receive a diagnosis in seconds. My design met my goal of creating a model which could output a diagnosis based off of just a skin lesion image. I also created a user-friendly website that makes the process of receiving a diagnosis easy and efficient. Going forward, I want to improve my Transfer Learning model as well as explore additional machine learning models. This way I can improve the diagnostic accuracy of my model. This is crucial as a false diagnosis in the medical field can be detrimental. By continuously improving my project, I hope to help those who struggle with skin cancer.