Skin Lesion Segmentation and Classification Using Conventional and Deep Learning Based Framework

Background: In medical image analysis, the diagnosis of skin lesions remains a challenging task. Skin lesion is common type cancer that exists worldwide. Dermoscopy one latest technologies used for cancer. Challenges: Many computerized methods have been introduced in literature to classify cancers. However, challenges remain such as imbalanced datasets, low contrast lesions, and extraction irrelevant or redundant features. Proposed Work: this study, new technique proposed based on conventional deep learning framework. The framework consists two major tasks: segmentation classification. task, initially improved by fusion filtering techniques then performed color transformation area discrimination. Subsequently, best channel selected map computed, which further converted into binary form using thresholding function. classification pre-trained CNN models were modified trained transfer learning. Deep features extracted from both fused canonical correlation analysis. During process, few also added, lowering accuracy. A called maximum entropy score-based selection (MESbS) solution issue. through approach are fed cubic support vector machine (C-SVM) final Results: experimental process was conducted datasets: ISIC 2017 HAM10000. dataset whereas HAM10000 achieved accuracy datasets 95.6% 96.7%, respectively, higher than existing techniques.