Dermoscopic Image-Based Deep Neural Network-Based Skin Lesion Diagnosis and Classification Model

Abstract

The article addresses a novel technique for identifying and classifying lesions in dermoscopic images. For lesion segmentation, this method uses a deep separable convolutional network. To start, black frames and extra hair must be removed to remove artificial and natural noise that could impede lesion localization. The photos are rotated and deformed for dataset augmentation after noise reduction. A segmentation model based on deep separable convolution and void convolution forms the basis of the process. To recreate the image's detailed features, features extracted during the encoding phase are fused during the decoding step. The proposed method is effective, as evidenced by the experimental findings, which yield a remarkable segmentation accuracy of 95.24%. This outperforms alternative segmentation models such as U-Net.