Measuring fairness in automatic skin disease diagnosis with a convolutional neural network

abstract

The aim of this project is to measure fairness in automatic skin disease diagnosis using a convolutional
neural network. The dataset used for this work is a compilation of various images of skin diseases taken
by the medical personnel from the pedriatic ward at Sant’Orsola hospital in Bologna. These images were
collected using consumer-grade cameras, resulting in considerable variance in illumination, angle, and
quality. The first part of this work involes processing the images to create a dataset with uniform image
sizes. The second part focuses on determining the skin tone of each sample using the Individual Typology
Angle. The third part involves training a simple convolutional neural network and evaluating its
performance across different skin tones by computing traditional metrics for each skin tone and different
fairness metrics.

keywords

Skin Tone Bias, Fairness, CNN, ITA

outcomes