Convolutional neural network-based approach for dermatological disease prediction

Abstract

The fast growing application of Artificial Intelligence in the field of medicine has led to an improvement in the detection and treatment of many kinds of ailments including skin diseases. However, there exists deficiency in the performances of the existing image pixel scaling techniques when applied to skin diseases detection. Pixel scaling is a major preprocessing tool in the classification of images. To improve the performance of skin disease detection and classification model, this paper proposes a new pixel scaling technique called Mean Pixel Division. At the preprocessing stage, each pixel value in the skin diseases images is divided by the mean of the entire channel pixel values. This reduces the range of the pixel values to a manageable level. A synthetic minority oversampling technique (SMOTE) was used to overcome the challenge of unbalanced classes’ distribution in the dataset. Then, a CNN architecture was designed and trained with the earlier pre-processed images. The evaluation of the proposed approach compared with some existing scaling techniques shows that our approach outperformed the existing techniques, having recorded 99.62%, 98.66% and 99.78% in terms of performance accuracy, sensitivity and specificity respectively