Skin cancer is the most prevalent diseases that can be identified visually and further using dermoscopic images. Visual observation gives an opportunity of utilizing computational intelligence techniques to intercept the different skin images, so various skin lesion classification methods using deep learning models show improved results. In this regard, this paper introduces a highly reliable and stable model for diagnosing skin diseases using deep learning architectures. Primarily, a preprocessing technique is proposed for eliminating noise and hair artifacts from the dermoscopy images. Secondarily, a modified U-Net is built for segmenting affected region from its surrounding tissues. Subsequently, two pretrained networks are separately employed for extracting deep features from the segmented lesion. Then, a light weight convolutional neural network is designed to classify images into seven classes against HAM10000 database. The presented lesion segmentation and classification methods’ outcomes are evaluated in light of accuracy, specificity, recall, precision, Mathews correlation coefficient, Jaccard index, F1-score and computational cost. Experimental outcomes show that the presented model provides excellent performance compared to other models with an accuracy of 99.67%.
Volume 12 | Issue 1
Pages: 485-498