Non-Hybrid Machine Learning Techniques for Classifying and Detecting Skin Disease Variants

Abstract

Eczema, acne, and psoriasis are all skin diseases that must be diagnosed early on to avoid complications. To detect and classify skin diseases, many researchers have developed a variety of support vector machine (SVM)-based classification models. However, these existing models suffer from imbalanced datasets, irrelevant feature selection, and difficulty in fine-tuning the SVM's hyperparameters. As a result, this study developed “Aquila Optimiser-Support Vector Machine (AO-SVM)” and “Harris Hawk Optimiser-Support Vector Machine (HHO-SVM)” to categorise eight (8) different skin diseases, “Granuloma Annulare (GRA)”, “Haemangioma (HEM)”, “Herpes (HEP)”, “Hidradenitis Suppurativa (HSP)”, “Keratocanthoma (KEC)”, “Lupus (LUP)”, “Sebaceous Hyperplasia (SEH)”, and “Sun Damaged Skin (SDS)”, using 2,700 photos of skin disease datasets, including 250 photos of each diseased dataset class and 700 photos of normal skin from the Kaggle village datasets.  The images were pre-processed, including reducing the size of the images, "digital hair removal using the Black-Hat transformation and inpainting algorithm", and eliminating noise, then the affected area was segmented using the Sobel edge detection method. The Grey Level Spatial Dependence and Colour Moment were then used to extract texture, shape, and colour features, and performance metrics such as false positive rate, specificity, accuracy, precision, and sensitivity were used to compare the efficiency of the two classification models (“AO-SVM” and “HHO-SVM”).  The results show that the “AO-SVM and HHO-SVM” classification models perform at 95.99% and 96.56%, respectively. This study adds to the body of knowledge by developing two refined Multiclass Support Vector Machine classification models, “AO-SVM and HHO-SVM”, for a subset of skin diseases. These models optimise the SVM classifier parameters (penalty cost, C, and kernel function, γ) to reduce false positives and improve classification accuracy.  In conclusion, these two models can be extremely useful in assisting people living in remote areas who have limited access to expert dermatologists in detecting their disease as soon as possible.