Equitable AI for Dermatology
AI is transforming healthcare, yet dermatology AI tools often underperform for people with darker skin tones due to a lack of diverse training data. This can lead to diagnostic errors, delayed treatments, and health disparities for underserved communities.
This challenge from Break Through Tech and the Algorithmic Justice League aims to help address this issue by building an inclusive machine learning model for dermatology. The first goal is to train a model that can classify 21 different skin conditions across diverse skin tones, using the datasets provided here in Kaggle as your starting point. Model performance is evaluated (based on a weighted average F1 score metric).
Through participating in this Kaggle competition, I’ve had the opportunity to learn more about the use of AI in the medical field, data processing for images, implementing various pre-trained models (Xception, Inception, GoogleNet, ResNet, and DenseNet), finetuning models, and implementing various data augmentation and sampling techniques to solve problems of limited and imbalanced data. I started with an initial F1a accuracy of 8% and increased it to 64% by the conclusion of the competition.