Skin Disease Classification Using Inception-ResNetV2 and Data Augmentation

Abstract

Skin plays an essential role in protecting the human body. Every year, various kinds of skin diseases affect the life quality of many people in the world. Despite the continuous improvement in clinical diagnosis, accurately classifying and diagnosing skin diseases require professional knowledge and remain to be a daunting task for doctors and patients. Many traditional machine learning algorithms, such as Nearest Neighbors and Support Vector Machines (SVM), have been used to solve this challenging task. But traditional machine learning requires manual feature extraction, which is inefficient and time-consuming. In recent years, the advent of Deep Learning and Convolutional Neural Networks (CNN) significantly improve the development of medical image classification. But little research and work in deep learning focus on skin disease classification with more than 10 types of skin diseases. To this end, I propose a Deep Learning method by employing Data Augmentation techniques to the Inception-ResNetV2 network architecture. Specifically, Data Augmentation is a strategy in diversifying the data available for the training model without collecting new data, and Inception-ResNetV2 is one of the mainstream deep learning architectures. The model is tested on the public Dermnet dataset, which consists of 23 types of skin diseases with approximately 19500 clinical images. Results on this dataset show that the proposed model reduces the overfitting problem and outperforms other mainstream deep learning architectures.