A Comparative Study on Automated Detection of Skin Diseases Using Various Image Segmentation and Deep Learning Approaches

This paper presents novel approaches to automatically detect skin diseases, namely acne, dermatitis, ringworm, skin cancer, and warts. A dataset was developed by combining skin disease images taken from reputable databases and from local patients. Due to the complex manifestations of image backgrounds, each segmentation technique including the YOLOv5 algorithm and the Simple Linear Iterative Clustering (SLIC) algorithm were applied to preprocess collected images into more accurate and simple image clusterings. Various pre-trained models including ResNet18, ResNet50, and EfficientNetB3 were studied in order to extract relevant features. Several recognition models were developed including Multilayer Perceptron models (MLP), Support Vector Machine models (SVM), and two pre-trained deep learning models VGG16 and MobileNetV1. A total of 76 different configurations were developed and tested. Experimental results show that among the different trained models, the deep learning model MobileNetV1 based on the SLIC technique led to the highest average accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 9. 4 2 \%}$</tex>. The recognition accuracy of each class of acne, dermatitis, ringworm, skin cancer, wart, and healthy skin was: <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 0 0 \%} \% \mathbf{9 8. 7 2 \%}, \mathbf{9 9. 8 1 \%}$</tex>, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 9. 4 5 \%}, \mathbf{9 9. 2 7 \%}$</tex>, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 9. 2 7 \%}$</tex>, respectively.