A deep learning based framework for Classifying the skin lesion in Dermoscopic images

Abstract

Skin cancer is among one of the most prevalent kinds of cancer, with around 2 to 3 newlinemillion incidences reported each year globally. Skin lesions are caused by an atypical cell newlineformation on the skin, and visual inspection of skin lesions is a challenging, newlinedifficult, time-prone, and complicated task. Computer Aided Diagnosis (also called newlineCAD) approaches can help practitioners improve their diagnostic abilities and decrease newlinethe time required to reach an accurate diagnosis. In addition, the lack of modern, easy to newlineuse CAD approaches has created severe concerns regarding the non-intrusive, accurate, newlinefast detection of diseases. CAD systems can aid in the early detection of skin lesions, newlineallowing patients to receive prompt treatment and increase their chances of survival. newlineHowever, considering the distinct and complicated features of skin lesion images, newlineevaluation of skin lesion images remains challenging. The presented work is to examine newlinevarious preprocessing, segmentation, and classification algorithms for analyzing skin newlinelesions in order to distinguish among cancerous vs non-cancerous images. newlineTo obtain accurate and reliable skin cancer detection, such challenges require the newlineutilization of sophisticated deep learning techniques and rigorous model development. newlineAccording to emerging research, in multi-classes skin cancer classification, convolutional newlineneural networks outperform and offering valuable support to dermatology professionals. In newlinethe proposed system, firstly a preprocessing technique has been developed that includes newlinehair and noise removal, data augmentation, and image resizing with model parameters. newlineThen training has been performed using transfer learning on the fusion of DenseNet121 newlineand ResNet50V2 model on the HAM10000 dataset. The proposed technique s newlineperformance is assessed using measures such as sensitivity, f-1 score, specificity, and newlineaccuracy.