A Hybrid Approach for Optimal Multi-Class Classification of Neglected Tropical Skin Diseases using Multi-Channel HOG Features

Abstract

Problem statemen: According to the WHO, the eradication of neglected tropical skin diseases is possible by 2030 if an early test to distinguish these diseases is available. For us, the analysis of the skin in the plaque or nodule phase could be an avenue to explore. Unfortunately, the most common and disabling diseases are often of bacterial origin and therefore have almost the same development and characteristics in their initial phase.Purpose of the study: Our goal is, therefore, to propose an efficient and simple method to provide an optimal dataset and an optimized multi-class classification method for early diagnosis.Method: The method consists of extracting the optimal Histogram of Oriented Gradient (HOG) features by browsing the images through different basic cell sizes (CS) of the Buruli Ulcer (BU), Leprosy, and Cutaneous Leishmaniasis skin lesion images. Then we obtain another dataset made of the averages of the different databases from these cell sizes. In order to solve the multi-class classification problem of the Support Vector Machine (SVM), we introduced an Error Correcting Output Code (ECOC) framework optimized by a hybrid metaheuristic algorithm to optimize the diagnosis of several diseases simultaneously.Result : A preliminary study of the different Cell Size 2*2, 4*4, 8*8, 16*16 allowed us to have 5 training databases, one of which is extracted from the other four by computing the averages of the HOG features of CS 2*2, CS4*4, CS8*8, CS16*16. This last multichannel database is the one that obtained the best results after the implementation of the hybrid whale optimization algorithm and shark smell optimization algorithm with Error Correcting Code (WOA-SSO-ECOC-SVM) on Matlab. We obtained 89% accuracy on the multi-channel dataset, 72% for CS4*4, 72% for CS8*8, and 72% for CS16*16.Conclusion: This study shows that it is possible to achieve an optimized multi-class skin NTD classification with good accuracy by optimally selecting the appropriate HOG characteristics.Implication: This result makes it possible to consider the development of mobile applications that allow, just by taking a picture of the lesion, to identify the diseases. This equipment could be used by front-line medical staff and community health workers. This could solve the effects of isolation and poverty in the fight against these diseases.