DilatedSkinNet: A feature fusion induced intelligent framework for skin lesion extraction

Abstract

Melanoma is considered one of the most fatal skin cancer. It is harmful to human life if not detected early. Early detection and proper diagnosis are highly crucial to reduce the fatality rate due to melanoma. Therefore, in this article, we have developed a unique encoder–decoder-based DilatedSkinNet framework with several folds of novelties. The designed encoder network sandwiches a series of Lesion Detail Extraction (LDE) blocks and max pooling layers, capturing multi-scale features with reduced spatial dimensions. Also, the proposed encoder framework can extract diverse lesion features at various levels. The designed bridge block with a fine feature aggregator module connects the encoder to the decoder network, for a smooth transition of significant details while maintaining spatial relationships among the pixels. The developed decoder network projects in-depth features into segmented masks, with reduced extraction of healthy skin regions. The developed DilatedSkinNet network is trained on the ISIC 2016 dataset while tested on ISIC 2016 and unseen dermoscopic images from benchmarked datasets including ISIC 2017, ISIC 2018, and PH². The robustness of the designed DilatedSkinNet model is validated by comparing the objective measures, including accuracy, sensitivity, specificity, Dice Coefficient, and Jaccard Index, against 70 existing approaches. Furthermore, the efficacy of the developed DilatedSkinNet framework is corroborated using visual demonstration. Extensive experiments show that the designed DilatedSkinNet model shows its superiority compared to state-of-the-art methods and attains better performance in an unseen setup.