Early Skin Disease Detection Using Hybrid Machine Learning Model for Vision 2030 of Kingdom of Saudi Arabia

Abstract

Healthcare has been emphasized as vital to the Vision 2030 plan of the Kingdom of Saudi Arabia (KSA). The principal objective is to give easy access to advanced healthcare, including remote healthcare facilities. Early diagnosis of diseases plays a significant role in achieving this objective. This paper introduces an effective detection model for prior skin disease detection and avoids prolonging risks. Recent methods for skin disease detection used ML and DL techniques. However, bias, ineffective feature learning, over-fitting, high computations, and training time are significant problems in current ML and DL methods. Therefore, this paper proposes an advanced hybrid approach utilizing ML and DL-based feature learning for early skin disease detection. Image quality enhancement and digital hair removal are performed initially in the pre-processing stage. Skin lesion regions are segmented using the Optimized GrabCut-Hidden Markov Model (OGCHMM). The extraction of color, shape, and size features is done using Gray Level Co-occurrence Matrix (GLCM) and Histogram-of-Gradients (HOG) feature extractors and deep features using a pre-trained ResNet model. Finally, the skin disease images are classified using a Multi-class Transductive Hybrid Kernel Support Vector Machine (MCTHKSVM) classifier whose hyper-parameters are optimized using Golden Eagle Optimization (GEO) to reduce training time and improve detection accuracy. Experiments are conducted using the benchmark skin disease datasets. GEO-MCTHKSVM model attained accuracies of 98.78%, 98.87%, 99.13%, 96.5% and 99.04% for HAM10000, ISIC-2018, ISIC-2019, Dermnet and PAD-UFES-20 datasets, respectively with minimized model complexity, ensuring its effectiveness for utilization in advanced healthcare services.