U-Net-based segmentation and deep learning driven feature fusion for skin lesion classification in dermoscopic images

Abstract Clinical decision-making depends extensively on the development of machine-assisted technologies for the automatic processing of skin lesion images. Such tools aid experts in the early diagnosis of melanoma through dermatoscopic images. Accurate identification of melanoma skin lesion from RGB images is still one of the challenging task in computer vision. This research proposes a unique Computer-Aided Diagnosis (CAD) system based on deep learning and machine learning approaches for the early stage melanoma identification. The proposed approach uses full convolutional networks (FCNs) for feature extraction and an improved U-Net architecture for segmentation. Extra layers for enhanced boundary refinement help to introduce an upgraded 42-layer U-Net model to achieve accurate lesion delineation. Using numerous FCNs obtaining the most important characteristics of the lesion color, texture, and shape improves diagnosis precision. In order to obtain precise lesion delineation, an enhanced 42-layer U-Net model is presented, which incorporates additional layers for greater border refinement. Several FCNs are used to improve diagnosis accuracy by obtaining the most important features of the lesion, such as its color, texture, and shape. These features are further classified using machine learning techniques including Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Naïve Bayes (NB). Using the ISBI 2016 ISIC skin lesion dataset, the proposed method accomplishes a remarkable classification accuracy that improves the automated diagnosis of melanoma. Enhanced segmentation in conjunction with excellent feature extraction results in improved classification performance. The suggested method performs exceptionally well on the ISBI 2016 ISIC skin lesion dataset, dramatically increasing automated melanoma diagnosis.