Improving Mobile-Based Early Skin Disease Diagnosis for Melanin-Rich Skins
Keywords: CNN, Melanin-Rich Skins, Skins of Colour, Skin Disease, CAD, Transfer Learning, Improving Colored Skin Disease Diagnosis, Skins of Colour
TL;DR: This paper explores an approach to mobile-based skin disease diagnosis that aims at improving performance for skins of color particularly melanin-rich skins.
Abstract: The use of Computer Vision in skin disease diagnosis has become widespread, particularly with the development of Convolutional Neural Networks (CNNs) in recent years. However, one significant issue persists - the under-representation of skins of colour in datasets, resulting in biased models. This paper presents the design and development of a mobile-based application specifically aimed at detecting skin diseases in individuals with dark skin tones. To achieve this, transfer learning on a pre-trained CNN model (ResNet50v2) was utilized for classification, where the model was trained on a diverse dataset obtained from various sources. The dataset was filtered, augmented, and preprocessed to ensure representation of coloured skins. A comprehensive preprocessing pipeline was developed to improve performance on melanin-rich skins, while maintaining the model's robustness for lighter skin tones. Four models were trained for different body parts to narrow down the search and each of the model achieved over 0.8 F1-Score. The mobile application was purposefully developed to allow easy accessibility, enabling early diagnosis of skin diseases by empowering both civilians to monitor their skin health at home and practitioners to enhance their diagnostic capabilities.
Submission Category: Machine learning algorithms
Submission Number: 52
Loading